Hybrid Quantum AI Matches 96% Accuracy in Early Cancer Detection

Qubits to Cures: Quantum ML for Cancer Detection With Quantum Computing India — Ai4

Ai4

Hybrid quantum algorithms halve training data needs and accelerate convergence for tumor screening, but physical qubit noise keeps clinical deployment in exploratory phases.

AI-assisted summary of Qubits to Cures: Quantum ML for Cancer Detection With Quantum Computing India — original source. Check the original for full context.

Source published: 2025-08-28

Detailed notes

## Background Published by Ai4 on YouTube on August 28, 2025, this talk features Shadab Hussain, Co-Founder of Quantum Computing India. The presentation description highlights how Quantum Machine Learning (QML) leverages qubit computation to process complex medical imaging data with enhanced speed and accuracy, addressing the bottlenecks of traditional models to advance early cancer detection and precision medicine.

## TL;DR Shadab Hussain examines the urgent need for early cancer detection and explains how hybrid Quantum Machine Learning (QML) models can overcome classical deep learning limitations in high-dimensional imaging. He presents two IEEE-published studies applying hybrid classical-quantum neural networks to brain tumor MRI scans and breast histopathology biopsies, achieving 95% to 96% accuracy alongside notable data efficiency and rapid training convergence. However, Hussain underscores that QML remains largely in an exploratory proof-of-concept stage due to hardware noise, limited logical qubits, data encoding bottlenecks, and optimization hurdles like barren plateaus. The session concludes with a call for close interdisciplinary collaboration among physicists, data scientists, and healthcare professionals to translate quantum computing into clinical practice.

## Atmosphere & style The presentation opens on a sobering, deeply personal note regarding the human cost of late cancer diagnoses before pivoting into a structured and highly technical walkthrough. Hussain delivers his insights with the pragmatic clarity of an active data engineer and researcher, balancing enthusiasm for quantum potential with refreshing candor about the current developmental limits of the technology. The tone is academic, measured, and focused on grounded empirical benchmarks rather than speculative quantum hype.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=4s) Early cancer detection remains critically challenged by classical deep learning models that often miss subtle morphological indicators in high-dimensional imaging scans. - [03:08](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=188s) Classical neural networks struggle with generalization across unseen patient cohorts, frequently showing error margins between 3% and 10% in early diagnostic settings. - [03:08](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=188s) Quantum computing addresses classical high-dimensional bottlenecks by utilizing superposition and entanglement to represent exponentially vast Hilbert spaces with fewer computational units. - [05:32](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=332s) Practical QML operates as a hybrid architecture, passing classical feature embeddings into parameterized quantum circuits rather than replacing classical computing entirely. - [08:47](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=527s) A hybrid ResNet-50 and Quantum Neural Network pipeline achieved 96% accuracy on binary brain tumor detection, though accuracy decreased to 89% on multi-class classification. - [12:49](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=769s) Breast cancer histopathology testing showed exceptional quantum data efficiency, achieving 93.5% validation accuracy using only 50% of the training dataset and converging in 25 epochs compared to 40+ classically. - [12:49](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=769s) Angular encoding provides an effective mechanism for mapping dense, high-dimensional pathological image features into bounded quantum rotation angles. - [15:37](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=937s) Clinical deployment of QML is currently blocked by physical hardware noise, decoherence, and a severe shortage of fault-tolerant logical qubits. - [17:16](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=1036s) Major algorithmic bottlenecks in parameterized quantum circuits include barren plateaus during gradient optimization and the difficulty of selecting optimal data-encoding strategies. - [18:25](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=1105s) Realizing practical quantum advantages in medicine will require pairing quantum computing with Generative AI and fostering deep interdisciplinary collaboration between physicists, clinicians, and data scientists.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=4s) Introduction and the Urgency of Early Cancer Detection

Shadab Hussain introduces himself as the co-founder of Quantum Computing India and a lead engineer focused on data engineering, who is also authoring several books. Moving quickly past personal introductions, he poses a fundamental question to the audience: what if cancer could be reliably detected before clinical symptoms even appear? Delayed diagnosis drastically reduces treatment success rates. Hussain shares that he has seen this tragedy firsthand with friends whose parents received late-stage cancer diagnoses, leaving physicians unable to offer curative treatment, telling families that "you can just pray for their survival till whatever time they're alive." While artificial intelligence is advancing rapidly, early detection remains difficult because traditional deep learning models frequently miss subtle morphological indicators in high-dimensional imaging data.

To illustrate the sheer scale of the crisis, Hussain presents 2025 epidemiological projections from the American Cancer Society. In the United States alone, approximately 618,000 individuals receive new cancer diagnoses annually, and around 204,000 deaths are expected. Focusing on high-mortality malignancies, he highlights brain tumors and breast cancer as critical areas where early intervention is paramount. When detection is delayed, effective therapeutic windows close, making timely diagnostic screening a matter of life and death.

- Over 93,000 Americans are diagnosed with brain tumors annually, facing a 5-year survival rate of approximately 35.7%. - One in eight women (13.1%) will be diagnosed with invasive breast cancer during their lifetime. - Every minute cancer detection is delayed directly contributes to preventable mortality.

[03:08](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=188s) Limitations of Classical AI and the Quantum Computing Paradigm

Current diagnostic workflows rely heavily on classical computation, leveraging GPUs to run deep learning architectures alongside multi-step magnetic resonance imaging (MRI) enhancement and segmentation pipelines. However, classical methods face systemic bottlenecks. Medical scans consist of massive, high-dimensional datasets where early-stage pathological signatures appear only as minuscule, ambiguous dots. Classical neural networks struggle to capture these delicate features consistently, often taking excessive compute time to train. Furthermore, classical models suffer from poor generalization when exposed to unseen patient cohorts, routinely exhibiting a 3% to 10% error margin and lacking the robustness required for mission-critical early diagnostic reliance.

Quantum computing presents an alternative computational model by replacing classical binary bits with quantum bits, or qubits. While a classical bit is constrained strictly to a state of 0 or 1, a qubit exists in a continuous linear combination of states until measured. Quantum computation relies on two foundational principles of quantum mechanics:

- **Superposition**: Analogous to a coin spinning in the air, a qubit remains simultaneously in a combination of states and is not fixed as heads or tails until measurement forces a collapse. - **Entanglement**: A non-local correlation where the state of one particle directly influences the state of another across space, enabling interconnected multi-qubit systems to process complex relational data.

[05:32](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=332s) Foundations and Architecture of Quantum Machine Learning (QML)

Quantum Machine Learning operates at the confluence of quantum computing, classical machine learning, and quantum mathematical theory. Instead of replacing classical architectures outright, QML leverages quantum-mechanical processes to handle computational subroutines that scale poorly on classical hardware. Determining when to apply QML requires evaluating task complexity and data structure. Certain high-dimensional datasets that prove intractable for classical algorithms can be naturally mapped onto quantum states, unlocking novel diagnostic representations across healthcare, energy, aerospace, fintech, and biotechnology.

```mermaid flowchart LR accTitle: Hybrid Classical versus Quantum Machine Learning Architecture subgraph Classical["Classical ML Pipeline"] direction LR C_In["Classical Input Data"] --> C_Hidden["Hidden Layers / DNN"] --> C_Out["Deterministic Output"] end subgraph Quantum["Quantum ML Pipeline"] direction LR Q_In["Classical Feature Vector"] --> Q_Enc["Quantum Encoding / Qubit Mapping"] Q_Enc --> Q_Circ["Parameterized Quantum Circuit"] Q_Circ --> Q_Meas["Measurement & Optimization Loop"] Q_Meas --> Q_Prob["Probabilistic Outcome"] end ```

In a standard deep learning pipeline, data passes sequentially through fixed-dimension hidden layers to generate an output. By contrast, a hybrid quantum machine learning pipeline maps classical feature vectors into quantum state space via parameterized quantum circuits. These circuits utilize quantum gates to execute operations across entangled qubits before collapsing the state through measurement. Because quantum models are probabilistic, the measurement yields probability distributions that guide gradient optimization loops to iteratively refine circuit parameters for downstream classification tasks.

- QML exploits fundamental quantum properties including superposition, entanglement, model linearity, and measurement non-linearity. - Quantum circuits can represent exponentially vast Hilbert spaces using relatively few physical qubits. - Circuit architectures can be iteratively tuned against test validation data to achieve optimal convergence.

[08:47](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=527s) Research Case Study 1: Brain Tumor Classification with Hybrid QNNs

Hussain details his IEEE-published research developing a hybrid quantum-classical transfer learning framework for brain tumor classification using the open-source Brain Tumor Segmentation (BraTS) dataset. The pipeline extracts high-level spatial feature representations from MRI scans using a pre-trained ResNet-50 backbone. After applying dimensionality reduction, the extracted feature embeddings are mapped onto quantum states and fed into a parameterized Quantum Neural Network (QNN) for classification.

The team evaluated the architecture on both binary classification (tumor presence vs. non-tumor) and multi-class classification across three distinct brain tumor types alongside healthy controls. The binary classifier achieved exceptional diagnostic performance, reaching approximately 96% validation accuracy and a 0.96 area under the ROC curve (AUC-ROC), with 94% precision for non-tumor detection and high precision for tumor identification. When expanded to multi-class classification, overall accuracy dropped to approximately 89%, reflecting the current developmental frontier of QML compared to highly mature deep learning models.

- The hybrid ResNet-50 and QNN pipeline achieved 96% validation accuracy on binary brain tumor detection. - Multi-class classification across three tumor types achieved 89% accuracy with balanced precision and recall. - The experimental models were developed using PennyLane and Qiskit, and deployed on IBM Quantum hardware via Amazon Braket. - Quantum error correction was omitted in this phase and remains a designated focus for subsequent research iterations.

[12:49](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=769s) Research Case Study 2: Breast Cancer Detection and Data Efficiency

In a second IEEE-published study, Hussain and his team adapted their hybrid quantum transfer learning approach to breast histopathology classification using the BreaKHis dataset. The dataset comprises roughly 10,000 microscopic biopsy images split evenly between 5,000 malignant and 5,000 benign tissue samples. Features extracted via the ResNet family were compressed through dimensionality reduction to fit hardware qubit constraints and mapped into quantum Hilbert space using angular encoding.

The resulting quantum model achieved 95% classification accuracy, demonstrating remarkable data efficiency. While deep learning networks require massive training corpora to prevent overfitting, the quantum model retained high accuracy even when training data was drastically curtailed. Furthermore, the quantum network demonstrated superior convergence speed and computational economy compared to classical baselines.

- Angular encoding allowed efficient mapping of dense pathological image features into bounded quantum rotation angles. - The quantum layer exhibited robust resilience against noisy, high-dimensional input artifacts. - Training required only 25 epochs to surpass the performance benchmarks that took classical models 40 epochs to achieve. - The model showed substantial reductions in memory usage and energy consumption alongside fast inference speeds.

[15:37](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=937s) Current Challenges and Future Outlook in Quantum Machine Learning

Reflecting on the road ahead, Hussain emphasizes that QML remains in an exploratory research phase, with virtually no end-to-end production systems currently deployed in clinical environments. Progress is heavily constrained by hardware noise and environmental decoherence, making fault-tolerant quantum error correction the primary technological hurdle. Although quantum hardware manufacturers frequently announce processors boasting hundreds or thousands of physical qubits, the count of fault-tolerant *logical* qubits available for algorithm execution remains severely restricted.

Looking forward, Hussain identifies promising integration points between quantum computing and advanced natural language models. Merging quantum algorithms with generative AI (GenAI) architectures could drastically accelerate complex multi-modal clinical reasoning. Emerging algorithmic development is also expanding beyond neural architectures to include quantum counterparts of foundational machine learning algorithms, such as Quantum k-Nearest Neighbors (Q-kNN).

- Current hardware is limited by physical noise, compilation loss, and low logical-to-physical qubit ratios. - Scaling hybrid models requires mature quantum error-mitigation and error-correction frameworks. - Integrating QML with Generative AI and Large Language Models offers promising pathways for multi-modal medical diagnostics.

[17:16](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=1036s) Quantum ML Algorithms and Practical Implementation Challenges

Moving from classical paradigms to quantum-enhanced counterparts opens up a wide array of quantum machine learning (QML) algorithms. These include quantum K-nearest neighbors (QKNN), quantum support vector machines (QSVM), and various parameterized quantum circuits. However, implementing these models comes with serious foundational hurdles. A major obstacle lies in how classical training data is loaded and embedded into a quantum circuit. Practitioners frequently struggle with selecting the right quantum data encoding techniques. Without a rigorous grasp of underlying quantum principles, choosing an unsuitable encoding strategy or optimizer will inevitably lead to suboptimal model performance.

Another major theoretical and practical obstacle is the phenomenon of barren plateaus in optimization landscapes, which can severely hinder gradient-based training in parameterized quantum circuits. While academic researchers often promise dramatic theoretical speedups, real-world practicality tells a more grounded story: quantum models often struggle to match the high accuracy achieved by modern classical deep learning architectures. For now, much of the field remains experimental and theoretical. The foundational work is actively being established through proofs of concept (POCs) and exploratory research, alongside the standard privacy, security, and ethical concerns inherent across all artificial intelligence domains.

```mermaid flowchart TD accTitle: Practical Bottlenecks in Quantum Machine Learning subgraph DataPrep ["Data & Encoding Challenges"] A["Classical Data"] --> B["Quantum State Encoding"] B --> C{"Encoding Strategy Selection"} end

subgraph Optimization ["Circuit Training & Optimization"] C -->|Risk| D["Barren Plateaus"] C -->|Risk| E["Suboptimal Optimizers"] D --> F["Training Stagnation"] E --> F end

subgraph Evaluation ["Real-World Benchmark"] F --> G["Quantum POC Results"] H["Classical Deep Learning Benchmarks"] -.->|Comparison Gap| G end ```

- **Algorithm Lineup**: Includes Quantum K-Nearest Neighbors (QKNN), Quantum Support Vector Machines (QSVM), and parameterized variational quantum circuits. - **Key Technical Bottlenecks**: High-dimensional quantum state preparation/encoding, avoiding barren plateau optimization traps, and bridging the gap between theoretical speedups and classical deep learning benchmarks.

---

[18:25](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=1105s) Interdisciplinary Collaboration, Broader Industry Impact, and Closing

The presentation title "Qubits to Cures" serves not merely as a catchphrase, but as an urgent call for interdisciplinary collaboration. Unlocking meaningful quantum advantage in healthcare requires close synergy across quantum physicists, clinicians, and data scientists. Beyond oncology and medical diagnostics, quantum computing holds transformative potential across numerous other sectors wherever complex optimization challenges arise, including aerospace, e-commerce, telecommunications, and energy.

Furthermore, quantum computation plays a critical role in developing post-quantum cryptography (PQC) to defend against future quantum cyberattacks, as well as advancing computational chemistry and materials discovery. Major industry pioneers driving hardware and software advancements include D-Wave, Google, IonQ, and IBM Quantum, alongside a vibrant and growing broader ecosystem. In closing, attendees are invited to scan the on-screen QR code to download the presentation deck as the session transitions into audience Q&A.

- **Cross-Domain Use Cases**: Energy grid optimization, aerospace logistics, telecommunications routing, e-commerce recommendation engines, post-quantum cryptographic security, and molecular simulation for materials discovery. - **Notable Industry Leaders**: D-Wave, Google Quantum AI, IonQ, and IBM Quantum.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=4s) you can just pray for their survival till whatever time they're alive. — Shadab Hussain

> [00:04](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=4s) Every minute cancer detection is delayed directly contributes to preventable mortality. — Shadab Hussain

> [05:32](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=332s) Certain high-dimensional datasets that prove intractable for classical algorithms can be naturally mapped onto quantum states, unlocking novel diagnostic representations across healthcare, energy, aerospace, fintech, and biotechnology. — Shadab Hussain

> [05:32](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=332s) Quantum circuits can represent exponentially vast Hilbert spaces using relatively few physical qubits. — Shadab Hussain

> [12:49](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=769s) While deep learning networks require massive training corpora to prevent overfitting, the quantum model retained high accuracy even when training data was drastically curtailed. — Shadab Hussain

> [15:37](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=937s) Although quantum hardware manufacturers frequently announce processors boasting hundreds or thousands of physical qubits, the count of fault-tolerant logical qubits available for algorithm execution remains severely restricted. — Shadab Hussain

> [17:16](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=1036s) Without a rigorous grasp of underlying quantum principles, choosing an unsuitable encoding strategy or optimizer will inevitably lead to suboptimal model performance. — Shadab Hussain

> [17:16](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=1036s) While academic researchers often promise dramatic theoretical speedups, real-world practicality tells a more grounded story: quantum models often struggle to match the high accuracy achieved by modern classical deep learning architectures. — Shadab Hussain

> [18:25](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=1105s) The presentation title 'Qubits to Cures' serves not merely as a catchphrase, but as an urgent call for interdisciplinary collaboration. — Shadab Hussain

> [18:25](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=1105s) Unlocking meaningful quantum advantage in healthcare requires close synergy across quantum physicists, clinicians, and data scientists. — Shadab Hussain

## Mentioned Shadab Hussain, Quantum Computing India, Ai4, American Cancer Society, IEEE, Brain Tumor Segmentation (BraTS) dataset, BreaKHis dataset, ResNet-50, PennyLane, Qiskit, IBM Quantum, Amazon Braket, D-Wave, Google Quantum AI, IonQ

## Video description Shadab Hussain, Co-Founder, Quantum Computing India

Abstracts

Quantum Machine Learning (QML) is revolutionizing the intersection of artificial intelligence and healthcare, offering groundbreaking advancements in medical diagnostics. In this talk, we will explore how QML enhances cancer detection by leveraging the computational power of qubits to process complex medical imaging data with unprecedented speed and accuracy. Traditional machine learning models face challenges in handling high-dimensional healthcare data, but quantum-driven approaches open new possibilities for early detection and improved patient outcomes. Through real-world research and practical insights, this session will showcase how QML is shaping the future of precision medicine and transforming the way we diagnose and treat critical illnesses.

[Source](https://www.youtube.com/watch?v=UNwVUhm-DDk)

中文简报

混合量子AI在早期癌症检测中达到96%准确率混合量子算法将肿瘤筛查的训练数据需求减半并加快收敛速度，但物理量子比特噪声使临床部署仍处于探索阶段。 Shadab Hussain 探讨了早期癌症检测的迫切需求，并阐述了混合量子机器学习（QML）模型如何克服经典深度学习在高维成像中的局限性。他展示了两项发表于 IEEE 的研究，这些研究将经典-量子混合神经网络应用于脑肿瘤 MRI 扫描和乳腺组织病理活检，在达到 95% 至 96% 准确率的同时，表现出显著的数据效率与快速的训练收敛。然而，Hussain 强调，由于硬件噪声、有限的逻辑量子比特、数据编码瓶颈以及贫瘠高原（barren plateaus）等优化障碍，QML 在很大程度上仍处于概念验证的探索阶段。会议最后呼吁物理学家、数据科学家和医疗专业人员展开紧密的跨学科合作，以将量子计算转化为临床实践。传统深度学习在处理高维医学图像时容易遗漏早期微小病理特征，且在未见患者群体中常出现 3% 至 10% 的误差率；量子计算则利用叠加与纠缠特性在高维希尔伯特空间中高效表征复杂数据。 (3:08) 现阶段实用的量子机器学习（QML）多采用混合架构，将经典特征嵌入输入参数化量子线路，而非完全取代经典计算系统。 (5:32) 基于 ResNet-50 与参数化量子神经网络（QNN）的混合模型在脑肿瘤二分类任务中达到了 96% 的验证准确率，但在扩展至多类别分类时准确率降至 89%。 (8:47) 乳腺癌病理图像测试展示出显著的数据效率：混合量子模型仅使用 50% 的训练数据即可达到 93.5% 的准确率，并在 25 个 epoch 内完成收敛，优于经典模型所需的 40 多个 epoch。 (12:49) QML 在临床的大规模部署仍受限于物理硬件噪声、退相干现象以及容错逻辑量子比特的严重不足。 (15:37)

Reading summary

Quantum Computing India co-founder Shadab Hussain argues that hybrid quantum machine learning can resolve classical computing bottlenecks in early cancer detection, though physical hardware noise and optimization hurdles currently confine the technology to experimental research. Diagnostic Limitations of Classical Oncology AI (0:04) Delayed cancer diagnosis severely undermines curative treatment, a crisis illustrated by American Cancer Society 2025 projections citing approximately 618,000 new cancer diagnoses and 204,000 deaths annually in the United States. Hussain notes that high-mortality conditions demand faster intervention: over 93,000 Americans are diagnosed annually with brain tumors, which carry a 35.7% five-year survival rate, while one in eight women faces invasive breast cancer. Existing clinical AI workflows rely on graphics processing units (GPUs) running deep learning networks on complex magnetic resonance imaging (MRI) scans. However, classical networks struggle to detect subtle, early-stage morphological markers, suffer from long training times, and show poor generalization across new patient cohorts, frequently displaying error margins between 3% and 10%. To address these high-dimensional computational bottlenecks, quantum computing replaces binary bits with qubits that leverage superposition and entanglement to model intricate relational data. Hybrid QNN Case Studies in Oncology (8:47) Rather than discarding classical architectures, hybrid Quantum Machine Learning (QML) pairs classical deep learning with parameterized quantum circuits. In an IEEE-published study using the BraTS brain tumor dataset, Hussain's team extracted spatial features from MRI scans using a pre-trained ResNet-50 backbone, compressed the embeddings, and classified them via a parameterized Quantum Neural Network (QNN) deployed on IBM Quantum hardware via Amazon Braket. The pipeline achieved 96% validation accuracy and a 0.96 AUC-ROC on binary classification, though accuracy dropped to 89% when expanded to multi-class classification across three tumor types. A second study applied hybrid transfer learning to 10,000 breast histopathology images from the BreaKHis dataset. By using angular encoding to map image features into quantum Hilbert space, the quantum model reached 95% accuracy in 25 epochs, outperforming classical models that required 40 epochs. The quantum framework demonstrated high data efficiency, maintaining 93.5% accuracy even when training data was halved to 50%, alongside reduced memory and energy requirements. Hardware Decoherence and Algorithmic Bottlenecks (15:37) Despite theoretical computational advantages, Hussain emphasizes that QML has not achieved end-to-end clinical deployment. Physical hardware remains constrained by environmental decoherence, compilation losses, and an absence of fault-tolerant quantum error correction. While manufacturers produce chips with hundreds or thousands of physical qubits, the count of stable, fault-tolerant logical qubits remains severely restricted. Algorithmic challenges also impede practical adoption. Encoding dense, classical patient data into quantum circuits is technically complex, and selecting an incorrect encoding scheme or optimizer can trigger barren plateaus—flat optimization landscapes that arrest gradient descent training. Consequently, quantum prototypes often struggle to consistently outperform mature classical deep learning baselines in real-world environments. Interdisciplinary Integration and Cross-Sector Applications (18:25) Translating quantum research into medical applications requires sustained collaboration among quantum physicists, data scientists, and clinicians. Hussain highlights future integration pathways between QML and Generative AI or Large Language Models to support multi-modal clinical diagnostics, as well as broader implementations of non-neural quantum algorithms such as Quantum k-Nearest Neighbors (Q-kNN) and Quantum Support Vector Machines (QSVM). Beyond oncology, Hussain notes that industry players—including IBM Quantum, Google Quantum AI, IonQ, and D-Wave—are developing quantum solutions for complex optimization problems in energy grid management, aerospace logistics, telecommunications, and materials discovery, as well as post-quantum cryptography to counteract future cybersecurity threats. While hybrid quantum machine learning exhibits substantial data efficiency and diagnostic accuracy in benchmark studies, its real-world clinical viability depends on solving quantum state encoding bottlenecks, navigating barren plateaus, and transitioning from noisy hardware to fault-tolerant logical qubits.

精炼阅读摘要

Quantum Computing India联合创始人Shadab Hussain指出，早期癌症筛查长期受制于经典深度学习处理高维影像时的算力瓶颈与微小病灶漏诊风险；而混合量子机器学习（QML）在数据利用效率与收敛速度上展现出颠覆性潜力，但受制于当前硬件噪声与优化难题，该技术整体仍处于实验室概念验证阶段。早期筛查紧迫性与经典计算的系统性瓶颈 (0:04) 癌症早期确诊对于提高患者生存率具有决定性作用。根据美国癌症协会2025年的流行病学预测，美国每年新增癌症确诊病例约为61.8万例，预计死亡人数约20.4万例。以高致死率恶性肿瘤为例，美国每年有超过9.3万人确诊脑肿瘤，其5年生存率仅约为35.7%；同时，八分之一的女性（13.1%）在一生中会被诊断出患有浸润性乳腺癌。然而，临床早期病灶在复杂的磁共振成像（MRI）等高维医学扫描中往往仅表现为细微且形态模糊的微点，传统筛查手段极易错失最佳干预窗口。当前基于图形处理器（GPU）与深度学习的主流诊断管线在面对海量高维数据时暴露出明显的计算瓶颈。经典神经网络不仅训练耗时极长，而且在面对未见患者群体时泛化能力不足，常常出现3%至10%的误差幅度。这种鲁棒性的缺失，使得依赖传统计算模型进行关键早期诊断的可靠性受到严重制约。混合量子机器学习的机制与状态映射 (5:32) 量子计算通过引入量子比特（qubit）突破了经典二进制状态的限制。利用量子力学的两大核心特性——叠加态（使量子比特在测量前处于多种状态的连续线性组合中）与量子纠缠（实现多粒子间的非局域关联），量子系统能够以相对较少的物理比特表达指数级庞大的希尔伯特空间，从而在数学上天然适配复杂关系和极高维数据的表征任务。在实际工程架构上，混合量子机器学习（Hybrid QML）并非直接取代经典体系，而是构建协同管线。该架构首先利用经典卷积骨干网络提取图像特征向量，随后通过量子态编码将特征映射至参数化量子电路（PQC）。电路通过量子门在纠缠比特间执行运算，最终经由量子测量塌缩生成概率分布，并借助经典梯度优化循环迭代调整电路参数，以极高的空间压缩效率完成下游分类与推断任务。脑瘤与乳腺癌检测实证与极高数据效率 (8:47) Hussain团队展示了两项已发表于IEEE的实证研究成果。在基于BraTS数据集的脑肿瘤筛查实验中，研究团队结合预训练的ResNet-50骨干网络与参数化量子神经网络（QNN），并在IBM Quantum硬件及Amazon Braket平台上完成部署。实验结果显示，该混合模型在肿瘤二分类任务中实现了约96%的验证准确率与0.96的ROC曲线下面积（AUC-ROC），对非肿瘤检测的精确率达到94%；但在拓展至涵盖三种肿瘤类型的多分类任务时，受限于现有QML技术成熟度，整体准确率回落至约89%。在基于BreaKHis数据集（包含约10,000张显微活检图像，良性与恶性各占5,000张）的乳腺癌病理切片研究中，团队采用角度编码技术映射图像特征。实验表明，量子模型不仅取得了约95%的分类准确率，更展现出卓越的数据效率与收敛特性：当训练数据量被大幅压缩至75%与50%时，量子模型仍分别维持了94.8%和93.5%的高准确率，且仅需25个epoch即可完成收敛；相比之下，经典深度学习基线需要40个以上的epoch才能逼近类似水准，凸显出量子层对抗输入噪声和降低能耗的显著优势。算法优化陷阱与缺乏容错硬件的现实制约 (15:37) 尽管实验室验证展现了诱人前景，但QML向真实临床落地仍面临深层技术障碍。首先在数据与算法层面，将高维经典病理数据高效嵌入量子线路极为复杂，一旦状态编码策略或优化器选择不当，模型极易陷入参数化量子电路特有的“贫瘠高原”（Barren Plateaus）优化陷阱，导致梯度消失和训练停滞。在诸多真实场景的对比测试中，量子原型往往仍难以匹敌经过高度工程优化的现代经典深度学习模型。其次在物理硬件层面，环境退相干与硬件物理噪声构成了根本掣肘。目前工业界虽然推出了包含数百甚至数千物理比特的处理芯片，但能够执行高保真度容错算法的逻辑量子比特数量仍然极度匮乏。研究团队在前期实验中亦尚未引入量子纠错机制。Hussain坦言，目前全球医疗临床环境中尚无端到端投入生产运行的QML系统，技术整体仍处于探索与概念验证阶段。 Hussain总结指出，“从量子比特到治愈方案”不仅是技术愿景，更是横跨量子物理学、临床医学与数据科学的紧迫协同任务。当前混合量子机器学习在抗癌诊断中展现出高数据效率与快速收敛的理论优势，但真正确立临床可用性仍有赖于容错量子纠错技术的突破，以及未来与生成式AI（GenAI）和大语言模型的融合发展。

Detailed notes — zh

## Background 该视频由 Ai4 于 2025 年 8 月 28 日发布在 YouTube 平台，主讲人为 Quantum Computing India 联合创始人 Shadab Hussain。内容主要聚焦于量子机器学习（QML）在医疗诊断中的突破性应用，探讨如何借助量子计算突破传统 AI 在处理高维医学影像时的算力瓶颈，以推动癌症极早期检测与精准医疗的发展。 ## TL;DR 演讲首先剖析了传统深度学习在脑肿瘤和乳腺癌等高维医学影像分析中所面临的漏诊风险与计算瓶颈，随后阐释了基于量子叠加与纠缠态的量子机器学习（QML）基本原理。讲者重点介绍了团队发表于 IEEE 的两项研究成果：一是结合 ResNet-50 与量子神经网络（QNN）实现脑肿瘤的高精度二分类与多分类检测，二是在 BreakHis 乳腺癌数据集上验证了量子模型在极少训练样本下依然具备出色泛化力与极快收敛速度的计算优势。最后，演讲客观分析了量子纠错、逻辑量子比特稀缺及变分优化中的“贫瘠高原”等现实挑战，呼吁跨学科力量携手推动 QML 的实际落地。 ## Atmosphere & style 演讲开篇以身边亲友确诊癌症晚期的无力感切入，迅速奠定了关切生命与迫切寻求技术突破的情感基调。整体氛围兼具严谨的学术深度与务实的工程态度，讲者在展示量子算法优越指标的同时，坦诚指出当前模型在细粒度多分类和硬件噪声下的局限性，绝不夸大其词。整场报告条理分明、层层递进，既有前沿科学探索的锐气，又蕴含着面向真实世界临床痛点的深厚人文关怀。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=4s) 癌症早期筛查对提升患者生存率至关重要，但现有经典深度学习在识别 MRI 影像中微小且隐蔽的早期病灶时常存在漏诊与鲁棒性不足的痛点。 - [02:06](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=126s) 传统深度学习模型在医疗诊断中面临高维特征捕获困难、高质量标注数据稀缺、训练耗时长以及跨分布泛化能力受限等多重算力与算法瓶颈。 - [04:25](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=265s) 量子机器学习（QML）利用量子比特的叠加态与纠缠态，将经典医学数据映射至量子线路执行并行概率优化，开辟了处理高复杂度医疗数据的新路径。 - [08:47](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=527s) 结合经典 ResNet-50 特征降维与参数化量子神经网络（QNN）的混合架构，在 BraTS 脑肿瘤二分类检测中取得了 96% 的高准确率与 AUC 表现。 - [10:13](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=613s) 在细粒度的脑肿瘤多分类任务中 QML 准确率回落至 89% 左右，表明量子机器学习算法目前仍处于演进阶段，尚未全面超越成熟的经典深度学习。 - [12:50](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=770s) 在 BreakHis 乳腺癌病理研究中，QML 展现出极高的小样本学习效率：即便训练数据集砍减 50%，模型准确率依然维持在 93.5%。 - [12:50](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=770s) 量子混合模型具备更快的收敛速度（25 个 Epoch 优于经典模型 40 个 Epoch），并在内存占用、能耗及抗噪鲁棒性上展现出显著潜力。 - [15:37](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=937s) 环境噪声干扰、量子纠错（Error Correction）机制的缺失以及真正可用于容错计算的“逻辑量子比特”数量不足，是制约 QML 走向生产环境落地的核心硬件瓶颈。 - [17:08](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=1028s) 算法层面存在经典数据载入量子态的编码瓶颈，以及变分量子线路中梯度指数级衰减的“贫瘠高原（Barren Plateaus）”难题，需谨慎选择线路架构与优化器。 - [17:08](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=1028s) 量子计算在精准医疗、后量子密码、材料科学等关键领域的突破，亟需量子物理学家、数据科学家与临床医学专家展开深度的跨学科协同创新。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=4s) 讲者背景与癌症早期筛查的现实痛点讲者 Shadab Hussain 进行了简短的自我介绍。他是 Quantum Computing India 的联合创始人，同时在 MCCO 担任首席数据工程师，目前还撰写并正在推进几本技术著作。Shadab 直奔主题，聚焦于如何将量子机器学习（Quantum Machine Learning, QML）应用于医疗健康领域。他向现场观众抛出了一个直击核心的问题：我们是否能在癌症症状出现之前就检测出病灶，从而大幅提高治疗的成功率？ Shadab 分享了身边的真实故事——他有几位朋友的父母在确诊癌症时已经是晚期，医生对此束手无策，直言“我们无法治愈它，只能为他们还能活着的每一天祈祷”。他指出，尽管当前 AI 领域进展迅猛，但若能借助更先进的算法与量子算力进一步提升预测精度，在极早期阶段确诊重大疾病，患者的生存希望将大大增加。在癌症筛查中，每一分钟的延误都意味着生命的流逝。目前许多公司专注于构建深度学习模型，但在处理复杂的 MRI 图像增强与分割时，现存技术仍容易忽略微小的潜在病变模式（subtle patterns），且诊断流程冗长，延误了最佳治疗时机。 ### [02:06](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=126s) 癌症统计数据与传统深度学习的算力局限根据美国癌症协会（American Cancer Society）公布的 2025 年相关统计数据，癌症对人类健康构成了严峻威胁。仅针对脑肿瘤与乳腺癌这两种致命病症，数据尤为触目惊心：脑肿瘤患者的 5 年生存率仅约为 35.7%，全美预计有超过 9.3 万人确诊；而大约每 8 名女性中就有 1 人（约 13.1%）在一生中可能被诊断出浸润性乳腺癌。如果无法及时发现，患者就会错过最佳用药和治疗时机。 Shadab 分析了当下传统计算与 GPU 在医疗影像处理上面临的瓶颈： - **高维数据难以全面捕获**：医疗图像维度极高，模型训练时极易遗漏微小特征。在 MRI 图像上，早期的癌变往往只是极其微小的斑点，现有的经典 AI 模型很难直接且稳定地捕捉到它们。 - **训练耗时长且数据稀缺**：训练大型深度学习模型需要耗费大量算力与时间，且医疗领域高质量标注数据有限。 - **泛化与鲁棒性不足**：当引入全新分布的数据时，模型预测失误的概率通常在 3% 到 10% 左右。尽管部分深度学习模型已获得医疗审批，但其鲁棒性仍不足以让医生在早期筛查中对其完全信赖。 ### [04:25](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=265s) 量子计算基础与量子机器学习（QML）的运作范式为了突破经典计算的瓶颈，量子计算提供了全新的底层路径。经典计算机基于二进制比特（0 或 1），而量子计算机使用量子比特（Qubit）。在被测量之前，量子比特可以处于 0 和 1 之间的任意叠加态。Shadab 用通俗的例子解释了量子力学的两个核心特性： - **叠加态（Superposition）**：就像在空中旋转抛出的硬币，在落地测量前，无法断定它是正面还是反面。 - **纠缠态（Entanglement）**：可以理解为“任何行动都会产生联动反应”，在宇宙中扰动一个粒子，另一个处于纠缠状态的粒子也会受到影响。量子机器学习（QML）正处于量子计算、经典机器学习与量子力学数学理论的交叉点。它利用量子物理过程来优化传统计算机难以承受的高复杂度计算路径。Shadab 指出，QML 并非适用于所有场景，需要根据任务与数据类型进行划分：经典模型难以处理的数据集，往往更适合交由量子模型运算。目前该技术已广泛探索于能源、航空航天、金融科技、生物医药等领域。 ```mermaid flowchart TD accTitle: 经典机器学习与量子机器学习流程对比 subgraph Classical["经典机器学习 (Classical ML)"] A["输入数据 (Input Data)"] --> B["隐藏层特征提取 (Hidden Layers)"] B --> C["输出结果 (Output)"] end subgraph Quantum["量子机器学习 (Quantum ML)"] D["经典数据 (Classical Data)"] --> E["量子态编码 (Data Encoding)"] E --> F["量子线路执行 (Quantum Circuit)"] F --> G["量子测量 (Measurement)"] G --> H["概率性输出 (Probabilistic Outcome)"] H --> I["线路优化与评估 (Optimization Loop)"] I -.-> F end ``` 在架构上，经典神经网络通过输入层、多层隐藏层最终输出确定性结果；而量子机器学习则是将经典数据映射并编码到量子比特上，利用量子线路（包含量子门与测量操作）执行并行优化计算。由于量子模型本质上具有概率性，系统通过多次测量获得结果分布，再评估测试集表现并反复反向优化量子线路参数。 ### [08:47](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=527s) 基于迁移学习与量子神经网络的脑肿瘤检测研究 Shadab 介绍了其团队在 IEEE 上发表的一篇利用迁移学习与量子神经网络（Quantum Neural Networks, QNN）进行脑肿瘤分类的科研论文。该研究基于公开的脑肿瘤数据集 BraTS（Brain Tumor Segmentation）展开，具体实现路径与阶段成果如下： ```mermaid flowchart LR accTitle: 基于 ResNet-50 与 QNN 的脑肿瘤分类流程 A["BraTS 影像数据"] --> B["ResNet-50 特征提取"] B --> C["降维与量子编码"] C --> D["量子神经网络 (QNN)"] D --> E1["二分类: 有/无肿瘤"] D --> E2["多分类: 三类脑肿瘤类型"] ``` - **特征提取与降维**：从 ResNet 模型家族中选定 ResNet-50 进行数据预处理与高维特征提取，随后对特征进行降维，并将其映射编码为量子态数据。 - **量子权重训练**：提取经典特征权重后，接入参数化量子神经网络（QNN）进行训练与优化。 - **两阶段分类任务**： - 第一阶段实现了二分类任务，用于快速判别输入影像中“是否存在脑肿瘤”。 - 第二阶段拓展至多分类任务，对脑部常见的 3 种不同类型肿瘤进行精准细分。 ### [10:13](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=613s) 二分类与多分类肿瘤检测模型的性能评估在图像肿瘤预测模型的评估中，研究团队将其与经典深度学习模型进行了详尽的对比分析。在二分类任务（判断是否存在肿瘤）中，量子机器学习模型展现出了非常出色的表现，验证集上的准确率（Accuracy）达到了约 96%。混淆矩阵显示，除了极少数假阴性（False Negative）和假阳性（False Positive）外，绝大多数预测都非常精准，ROC 曲线下面积（AUC）同样达到了 96%。具体指标方面，无肿瘤类别（标签 0）的检测精确率（Precision）为 94%，有肿瘤类别（标签 1）也维持了高水准的检测精度，证实了所设计量子算法的有效性。 ```mermaid flowchart LR accTitle: Quantum Image Classification Pipeline subgraph Data["数据与经典特征提取"] A["医学影像数据"] --> B["ResNet 骨干网络"] B --> C["特征降维"] end subgraph Quantum["量子处理部分"] C --> D["角度编码 (Angle Encoding)"] D --> E["变分量子线路 (VQC)"] E --> F["量子测量与优化"] end subgraph Output["分类结果"] F --> G["二分类 / 多分类预测"] end ``` 然而，当任务扩展到多分类（Multiclass Classification）场景时，模型的准确率出现了一定程度的下滑，整体准确率约为 89%。针对包含不同肿瘤亚型及正常组织的 0、1、2、3 各个类别，精确率与召回率（Recall）虽处于可用水平，但并不算特别突出。演讲者坦诚地指出，量子机器学习（QML）目前仍处于算法演进阶段，尚未达到全面超越经典深度学习模型的水平，在复杂的细粒度多分类任务中仍可能面临精度瓶颈。在此项实验中，团队主要基于 PennyLane 与 Qiskit 框架构建量子线路，并通过 AWS Braket 平台接入 IBM Quantum 真实量子计算机进行运算，目前阶段尚未引入量子纠错机制。 - **二分类准确率与 AUC**：均达到约 96%，无肿瘤样本（类别 0）精确率达到 94%。 - **多分类准确率**：整体准确率回落至 89% 左右。 - **开发工具链与硬件**：使用 PennyLane 和 Qiskit 编写算法，通过 AWS Braket 调度 IBM Quantum 硬件资源执行计算。 --- ### [12:50](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=770s) 基于 IEEE 发表成果的样本效率与计算优势验证随后，演讲者介绍了发表于 IEEE 的第二项研究成果。该研究采用了公开的乳腺癌病理数据集（BreakHis），共包含约 10,000 张样本图像（5,000 张癌变图像与 5,000 张非癌变图像）。受限于当前量子硬件的可用量子比特数，研究首先借助 ResNet 骨干网络提取图像特征并进行降维，接着通过角度编码（Angle Encoding）将经典特征映射至量子态，再输入优化后的量子线路进行训练与全面评估。实验结果表明，该量子混合模型的基准准确率稳定在 95% 左右。演讲者强调，“寻找最优的量子线路结构以及匹配适宜的优化器至关重要”，否则量子模型很容易因线路未优化而出现精度断崖式下跌。此外，该研究重点验证了量子计算在小样本学习上的泛化能力：当把训练数据集缩减至 75% 时，模型准确率依然保持在 94.8%；即便进一步将数据量砍减至 50%，准确率仍高达 93.5%。这表明量子模型无需海量数据即可完成泛化训练。在计算效率方面，量子模型仅需 25 个 Epoch 的训练便超越了经典模型 40 个 Epoch 的收敛效果，显著缩短了训练时间，并在内存占用、能耗及推理鲁棒性上展现出对抗高维和噪声输入的独特潜力。 - **数据集规模**：BreakHis 数据集约 10,000 张图像（5,000 张恶性，5,000 张良性）。 - **数据缩减测试**：使用 75% 数据时准确率为 94.8%，缩减至 50% 数据时准确率仍达 93.5%。 - **收敛速度对比**：量子模型 25 个 Epoch 的表现优于经典模型 40 个 Epoch。 - **核心优势**：对高维、带噪数据的鲁棒性显著增强，具备出色的样本利用效率。 --- ### [15:37](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=937s) 量子机器学习的前沿挑战与大模型融合前景谈及量子机器学习的未来展望，演讲者指出该领域仍处于非常早期的研究阶段，目前业界几乎没有实现大规模生产环境（Production）落地的案例，主要成果仍集中于顶尖期刊与专利探索。制约 QML 走向实用的首要障碍是量子纠错（Error Correction）：当前的量子计算机极易受到环境热噪声、退相干以及编译损耗的干扰，如何在算法层面对抗噪声是亟待解决的课题。另一大核心瓶颈在于物理量子比特与逻辑量子比特（Logical Qubits）的差距。尽管各大硬件厂商频繁宣布推出包含数百甚至上千量子比特的芯片，但真正能够用于容错计算的逻辑量子比特数量仍然极其有限，这直接限制了量子线路的深度和数据编码的维度。与此同时，前沿方向正探索将量子计算与大语言模型（LLM）、生成式 AI（GenAI）及自然语言处理相结合，期望借助量子加速来攻克某些特定复杂结构下的语义建模与推理任务。 - **量子纠错**：需解决环境噪声与编译过程中的误差累积。 - **量子比特质量**：关注可用逻辑量子比特数量，而非单纯宣传的物理比特总数。 - **AI 融合新方向**：量子计算与 LLM / GenAI 结合加速复杂问题求解。 --- ### [17:08](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=1028s) 算法理论瓶颈、贫瘠高原与跨学科落地呼吁在算法生态方面，经典机器学习算法在量子领域大多存在对应映射，例如量子 K-近邻（QKNN）和量子支持向量机（QSVM）等。然而，实际开发中面临着严峻的技术陷阱：首先是经典数据如何高效载入量子态的数据编码（Data Encoding）问题，若缺乏对量子物理原理的深入理解，不当的编码与优化器选择将导致性能崩溃；其次是量子变分线路中普遍存在的“贫瘠高原”（Barren Plateaus）现象，即梯度随量子比特数增加呈指数级衰减，导致模型无法有效训练。演讲者坦承，“理论上研究人员承诺了许多美妙前景，但在实际应用中，有时依然难以超越成熟的经典深度学习模型”，目前行业正在通过概念验证（PoC）夯实理论与工程基础，同时数据隐私、安全与 AI 伦理也是必须考量的维度。演讲最后强调，本次分享不仅是一次技术报告，更是面向跨学科协作的呼吁。攻克重大疾病检测等关键课题需要量子物理学家、临床医生和数据科学家紧密合作。除了医疗健康领域，量子计算在航天航空、电商推荐、能源调度、材料发现、后量子密码学（PQC）以及各类复杂的组合优化问题中均蕴含着巨大价值。包括 D-Wave、Google、IonQ、IBM Quantum 在内的众多行业领军企业正在积极推动软硬件发展，量子计算赋能产业的蓝图正逐步铺开。 - **常见 QML 算法**：QKNN（量子 K-近邻）、QSVM（量子支持向量机）等。 - **主要理论瓶颈**：数据编码瓶颈、变分优化中的贫瘠高原（Barren Plateaus）梯度消失问题。 - **跨领域应用场景**：医疗诊断、后量子密码（PQC）、材料科学、组合优化、航天与能源。 - **核心生态厂商**：IBM Quantum、Google、IonQ、D-Wave 等。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=4s) 我们是否能在癌症症状出现之前就检测出病灶，从而大幅提高治疗的成功率？ — Shadab Hussain > [00:04](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=4s) 我们无法治愈它，只能为他们还能活着的每一天祈祷 — Shadab Hussain > [00:04](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=4s) 在癌症筛查中，每一分钟的延误都意味着生命的流逝。 — Shadab Hussain > [04:25](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=265s) 经典模型难以处理的数据集，往往更适合交由量子模型运算。 — Shadab Hussain > [04:25](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=265s) 任何行动都会产生联动反应 — Shadab Hussain > [10:13](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=613s) 量子机器学习（QML）目前仍处于算法演进阶段，尚未达到全面超越经典深度学习模型的水平 — Shadab Hussain > [12:50](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=770s) 寻找最优的量子线路结构以及匹配适宜的优化器至关重要 — Shadab Hussain > [12:50](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=770s) 量子模型无需海量数据即可完成泛化训练。 — Shadab Hussain > [15:37](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=937s) 目前业界几乎没有实现大规模生产环境（Production）落地的案例，主要成果仍集中于顶尖期刊与专利探索。 — Shadab Hussain > [17:08](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=1028s) 理论上研究人员承诺了许多美妙前景，但在实际应用中，有时依然难以超越成熟的经典深度学习模型 — Shadab Hussain > [17:08](https://www.youtube.com/watch?v=UNwVUhm-DDk&t=1028s) 本次分享不仅是一次技术报告，更是面向跨学科协作的呼吁。 — Shadab Hussain ## Mentioned Shadab Hussain, Quantum Computing India, MCCO, Ai4, 美国癌症协会 (American Cancer Society), IEEE, BraTS 数据集, BreakHis 数据集, ResNet-50, PennyLane, Qiskit, AWS Braket, IBM Quantum, Google, IonQ, D-Wave ## Video description Shadab Hussain, Co-Founder, Quantum Computing India Abstracts Quantum Machine Learning (QML) is revolutionizing the intersection of artificial intelligence and healthcare, offering groundbreaking advancements in medical diagnostics. In this talk, we will explore how QML enhances cancer detection by leveraging the computational power of qubits to process complex medical imaging data with unprecedented speed and accuracy. Traditional machine learning models face challenges in handling high-dimensional healthcare data, but quantum-driven approaches open new possibilities for early detection and improved patient outcomes. Through real-world research and practical insights, this session will showcase how QML is shaping the future of precision medicine and transforming the way we diagnose and treat critical illnesses. [Source](https://www.youtube.com/watch?v=UNwVUhm-DDk)

Source quotes

[0:04] Shadab Hussain: 我们是否能在癌症症状出现之前就检测出病灶，从而大幅提高治疗的成功率？ [0:04] Shadab Hussain: 我们无法治愈它，只能为他们还能活着的每一天祈祷 [0:04] Shadab Hussain: 在癌症筛查中，每一分钟的延误都意味着生命的流逝。 [4:25] Shadab Hussain: 经典模型难以处理的数据集，往往更适合交由量子模型运算。 [4:25] Shadab Hussain: 任何行动都会产生联动反应 [10:13] Shadab Hussain: 量子机器学习（QML）目前仍处于算法演进阶段，尚未达到全面超越经典深度学习模型的水平 [12:50] Shadab Hussain: 寻找最优的量子线路结构以及匹配适宜的优化器至关重要 [12:50] Shadab Hussain: 量子模型无需海量数据即可完成泛化训练。 [15:37] Shadab Hussain: 目前业界几乎没有实现大规模生产环境（Production）落地的案例，主要成果仍集中于顶尖期刊与专利探索。 [17:08] Shadab Hussain: 理论上研究人员承诺了许多美妙前景，但在实际应用中，有时依然难以超越成熟的经典深度学习模型 [17:08] Shadab Hussain: 本次分享不仅是一次技术报告，更是面向跨学科协作的呼吁。

Source quotes

[0:04] Shadab Hussain: you can just pray for their survival till whatever time they're alive. [0:04] Shadab Hussain: Every minute cancer detection is delayed directly contributes to preventable mortality. [5:32] Shadab Hussain: Certain high-dimensional datasets that prove intractable for classical algorithms can be naturally mapped onto quantum states, unlocking novel diagnostic representations across healthcare, energy, aerospace, fintech, and biotechnology. [5:32] Shadab Hussain: Quantum circuits can represent exponentially vast Hilbert spaces using relatively few physical qubits. [12:49] Shadab Hussain: While deep learning networks require massive training corpora to prevent overfitting, the quantum model retained high accuracy even when training data was drastically curtailed. [15:37] Shadab Hussain: Although quantum hardware manufacturers frequently announce processors boasting hundreds or thousands of physical qubits, the count of fault-tolerant logical qubits available for algorithm execution remains severely restricted. [17:16] Shadab Hussain: Without a rigorous grasp of underlying quantum principles, choosing an unsuitable encoding strategy or optimizer will inevitably lead to suboptimal model performance. [17:16] Shadab Hussain: While academic researchers often promise dramatic theoretical speedups, real-world practicality tells a more grounded story: quantum models often struggle to match the high accuracy achieved by modern classical deep learning architectures. [18:25] Shadab Hussain: The presentation title 'Qubits to Cures' serves not merely as a catchphrase, but as an urgent call for interdisciplinary collaboration. [18:25] Shadab Hussain: Unlocking meaningful quantum advantage in healthcare requires close synergy across quantum physicists, clinicians, and data scientists.

Original transcript

[0:04] I'm Shadab and uh I'm still learning how [0:07] to operate this. Yeah. So uh basically [0:09] I'm the co-founder for quantum computing [0:11] India and I work as a lead engineer at [0:13] MCCO focusing on data enginei site. I [0:15] have authored a couple of book couple of [0:17] books I'm working on. So that's a bit of [0:20] intro about me. I will not waste your [0:22] time here. So uh today I'm going to talk [0:24] about uh like how we can apply quantum [0:26] machine learning for the healthcare. [0:29] First question I have with you like uh [0:31] have you ever wondered like uh whatever [0:34] like uh if we can detect cancer before [0:37] uh symptoms appear so that we can [0:40] increase the success rates for the of [0:42] the treatment because a lot of time like [0:44] uh it's my p uh like from my friends [0:47] also like a couple of friends I have uh [0:49] they in their family like either their [0:51] mom or dad they have I mean at the later [0:54] stage they got diagnosed with a cancer [0:57] and since it was at a very late stage [0:59] age the doctor is like we can't treat it [1:02] or cure it like we can you can just pray [1:04] for their survival till whatever time [1:06] they're alive. So the question is if we [1:09] can like in AI space there's a lot [1:11] happening but if we can enhance those [1:14] predictions with the help of other [1:16] algorithms or computing powers so that [1:20] such uh critical diseases can be [1:22] diagnosed at the early stage it can [1:24] increase the success rate for the [1:26] treatment and every minute if uh the [1:29] cancer detection is getting delayed [1:31] people are losing the lives and uh we [1:34] have a lot of like deep learning models [1:36] they companies solely focused on the [1:39] like building deep learning models using [1:41] the algorithms and all but there are [1:43] subal patterns which get missed and they [1:45] can't they can't detect the cancer at [1:47] the very early stage and there lot of [1:48] like combination of different techniques [1:49] like for enhancing the MRI images you [1:52] have to use different type of [1:54] segmentation techniques then you go for [1:57] like using any of the deep learning [1:58] models and it takes a lot of time and we [2:01] have other challenges as well uh which [2:03] delays the treatment or the diagnosis so [2:06] this is the current like cancer stats [2:08] like overall if you look at the numbers [2:11] uh and this is of 2025 from the American [2:14] Cancer Society so 618K are like the [2:17] patients which have been newly diagnosed [2:19] and they are like around 204k people who [2:23] are expected to die because of it. Then [2:25] if I just specifically focused on brain [2:27] tumor and breast cancer then the numbers [2:29] are little less but they are also the [2:31] deadliest cancer um uh which exist and [2:35] out of like approximately 35.7% [2:39] 5ear survival rate we have with 93k plus [2:42] Americans who are diagnosed with brain [2:45] tumors in 2025 for example and out of [2:49] like eight women every one like that is [2:52] like 13.1% will be diagnosed with [2:55] invasive uh breast cancer in their [2:56] lifetime and if you don't detect these [2:59] cancers like uh in time the cure is not [3:02] possible like they're too late in terms [3:04] of like getting treatment from the [3:06] doctors or getting the right medicines. [3:08] So uh even we have a lot of like uh GPUs [3:11] and classical computation power but we [3:13] have certain limitations and what those [3:15] limitations are the data on which we [3:18] train these models are of very high [3:20] dimension and it's not easy to capture [3:23] each and every piece of the information [3:25] from those images. So at times like even [3:27] when we train the model uh we miss out [3:30] on the some of the critical information [3:32] or to detect the cancer at the very [3:35] early stage because those are very small [3:37] dots if you look on the MRI images and [3:39] it's not possibly detectable with the [3:41] help of the AI models easily and when we [3:44] train those AI models they consume a lot [3:45] of time to train as well and the other [3:49] bigger problem is the lack of data. So [3:51] doesn't matter like how much healthcare [3:53] data exist still when we train the model [3:56] and when some new data comes in the [3:58] chances of model going wrong is there [4:00] like be it like 3% 5% or 10% approx we [4:04] take the hit on predicting the wrong [4:06] outcomes [4:08] and the robustness is not there still I [4:10] mean people there are like different [4:13] type of like approval for a lot of deep [4:14] learning models which people are using [4:16] but still it's not very robust enough [4:18] where we can rely completely on it for [4:22] the um earlier detections. [4:25] So that's where the quantum computing [4:27] comes in the picture and so like as we [4:31] have classical bits uh like z as we have [4:34] bits on the classical computer zero [4:36] zeros and one the same way on quantum [4:39] computers we have cubits. So cubits can [4:41] also have the state of zero and one but [4:44] it can be in between any of the states [4:46] till it gets measured. And there are a [4:48] couple of like quantum mechanics [4:50] properties on the basis of which the [4:52] whole quantum computers operate. So [4:54] those are superposition and [4:55] entanglement. So for those uh for those [4:58] of you like who are not aware of these [5:00] concept superposition is something where [5:02] for example if you're flipping a coin [5:04] until the coin is in the air we can't [5:07] say whether it is head or tail. Right? [5:09] So that is a superposition state and [5:12] entanglement is something like every [5:13] action has its own consequences you can [5:16] say like in the easy term. So if we [5:18] disrupt any particle in the universe, it [5:20] will have the effect on the other [5:21] particle. So that's what is uh [5:23] entanglement and the whole quantum [5:25] computers or there are different type of [5:27] quantum computers but they operate [5:28] mainly on these uh properties. [5:32] So going next um uh today I'm going to [5:35] cover more about the quantum machine [5:37] learning. So quantum machine learning is [5:38] at the intersection of u quantum [5:41] computing, machine learning and uh math [5:44] for quantum mechanics. So it takes the [5:47] advantage of quantum mechanical process [5:49] and induce the existing machine learning [5:51] algorithms as well and try to solve some [5:54] of the complex parts which can take a [5:56] lot of time to be solved on the [5:58] classical computers and try to optimize [6:00] that and uh run it and there are [6:03] different algorithms which are getting [6:05] developed. So we there we try to like [6:07] combine both of these and see like uh [6:09] what novelty we can get from the newly [6:12] derived algorithms there. [6:15] So this is like and we can't like just [6:17] go with the quantum machine learning use [6:19] cases with every like data or the task. [6:23] So we need to understand like what kind [6:24] of task and what kind of data we have [6:27] available and then we based on this [6:30] quadrant we are which we are looking at [6:32] we decide whether we should be like what [6:34] combination we should be taking up and [6:36] trying to like apply these techniques. [6:38] So a lot of times like we observe data [6:40] generated by the quantum circuit is hard [6:42] to simulate classically but if you if [6:45] you go for like data sets that are hard [6:46] for classical models at times they are [6:48] easy to operate on the quantum models. [6:52] U so there are different applications [6:54] for this like across the industry. So [6:57] beat like uh energy sector, aerospace, [7:00] uh fintech, network, medical, biotech [7:03] everywhere we have the applications of [7:05] quantum computers. uh but at the same [7:07] time like quantum machine learning is [7:09] exploiting unique properties and [7:11] behavior of the quantum systems. So uh [7:14] unity of nature mathematics and [7:15] computing indeterminism of quantum [7:18] systems uh superposition and integment [7:20] which we just talked about line uh [7:23] linearity of the quantum models [7:25] nonlinearity of the quantum measurement. [7:27] So these are some of the exploits unique [7:29] uh like how it exploits the unique [7:31] properties [7:33] and if I have to like explain you uh [7:36] what is the difference between classical [7:38] uh machine learning and quantum machine [7:39] learning. So it's a very simple uh [7:41] diagram over here on the left we have [7:43] the classical diagram where we take the [7:45] input and then we have the hidden [7:47] layers. For example, if you consider the [7:48] neural network whatever model we are [7:50] trying to train it will give us the [7:52] output after the model has been trained. [7:54] In case of quantum machine learning, we [7:57] will map the classical data to the [7:59] cubit. We will encode the data and there [8:02] are different kind of operations [8:03] operators which we use for that. And [8:05] then instead of having the hidden [8:07] layers, we will have circuits built [8:10] using cubits and classical bits as well [8:13] at times if needed depends upon like how [8:15] we are designing the algorithm. And then [8:16] the circuit execution will be like [8:18] happening which will optimize the whole [8:20] calculation path and then it we will [8:22] measure the outcome of it and quantum [8:24] models are probabilistic in nature. So [8:26] we'll get the probability of the outcome [8:28] and we can decide like what kind of like [8:30] outcome expectations we are looking for [8:32] and then we get the final outcome and [8:34] after getting the outcome we try to [8:36] evaluate the model on the test data and [8:37] then if needed we again optimize the [8:40] circuit and or if you're happy with the [8:42] result then we take that model and uh [8:44] use it further. [8:47] Um so uh we I have been researching on [8:50] quantum machine learning side a bit uh [8:52] uh for the healthcare. So this is one of [8:54] our paper which we published with ILE E [8:57] and uh there we tried detecting or you [9:01] can say classifying brain tumor and [9:03] using transfer learning and quantum [9:04] neural networks. So this is the paper [9:07] you can scan the QR code of this paper [9:09] if you want to read it. Uh at the end of [9:11] the talk I have the QR code for the [9:12] whole presentation. So feel free to just [9:14] listen to me right now and later on you [9:16] can just take the full presentation. So [9:18] this in this paper what we are trying to [9:20] like we have a data set and the data set [9:23] which we are using is brats and it has [9:27] different type of tumors data as well [9:29] classified. So we start with uh like we [9:32] had a whole family of resnet and out of [9:34] which we selected reset 50 and from that [9:38] after pre-processing we are extracting [9:40] the features and then we are mapping [9:41] those features after uh after doing the [9:44] dimensionality reduction to the uh [9:47] quantum data and then we extracting [9:50] those weights and then we are training a [9:52] quantum neural network from which the [9:54] first part of it like we are just trying [9:56] to classify the images whether that is [10:00] uh having the tumor or not. So that was [10:03] the first use case and then the second [10:04] use case we went ahead and we did [10:06] multiclassification. So we had three [10:08] types of tumor in brain and so we are [10:11] trying to classify what kind of tumor [10:13] that image is getting predicted with or [10:16] there's no tumor in that image and then [10:18] we are trying to evaluate the model [10:20] performance and comparing that with the [10:22] classical models and then sort of like [10:24] saving the results. So this is how uh [10:26] the result looked like. U so uh talking [10:30] in terms of like training and validation [10:32] accuracy we had approximately 96% of the [10:34] accuracy on the like validation data set [10:38] this is for the binary classification so [10:39] confusion metrics also you can look at [10:41] like we have like very accurate [10:42] predictions apart from like few false uh [10:45] negative and false positive here and [10:47] then ROC curve also like area under [10:49] curve a curve was also like 96% which [10:53] shows that quantum model was able to [10:54] like perform well with the algorithm [10:57] which designed. [10:59] So, and this is the like uh metrics on [11:02] the basis of which we tried evaluating [11:03] the model. So, accuracy was like around [11:05] 96% and um zero is like u non-tumor and [11:10] one is tumor here. So, 94% precision for [11:14] detecting uh non-tumor and one for like [11:16] detecting the tumors in the data set. [11:20] And uh this is the results for [11:22] multiclassification. So here we had a [11:24] dip in the like accuracy. You can see it [11:27] was not like very high accurate because [11:31] quantum machine learning is still not [11:33] there where it can beat the deep [11:35] learning models. U we are like people [11:38] are developing the algorithms we are [11:40] trying to get there but for all the use [11:42] cases of a multiclassification problems [11:44] at times it fails or it will have lower [11:46] accuracy or precisions in comparison to [11:48] the deep learning models and there are [11:50] multiple factors which contribute to the [11:52] lower accuracy which I will be coming uh [11:54] next but this was the like confusion [11:56] metrics and the how the uh loss function [11:59] looks like on the train and validation [12:01] data set. [12:02] So accuracy for this multiclassification [12:05] model was like around 89%. [12:08] And for each of the classes 0 1 2 3 like [12:11] tumors and like non-tumor one we have [12:14] precision and recall score also like at [12:16] a decent stage not very high here. So [12:20] that was the result for [12:21] multiclassification and um any question [12:24] so far like yeah error correction we [12:26] have not focused yet. So that's the like [12:28] uh feature scope of this work where we [12:31] are like enhancing it and trying to [12:33] adapt to other data set as well. Uh but [12:36] mainly we tried using penny lane and [12:37] tried and the cquskit framework and [12:39] running it on the IBM quantum computer [12:41] through AWS bracket platform. [12:45] So uh this was the result for uh [12:47] multiclassification [12:49] approach and this is our second work uh [12:52] which also got published in it e where [12:54] we followed the same approach [12:56] approximately with the different data [12:57] set. So we used break open-source data [13:01] set uh it has like approximately 10,000 [13:03] samples of 5,000 like cancerous images [13:06] and 5,000 non cancerous images. [13:10] Uh the approach was also similar uh here [13:12] where we started with um ResNet family [13:15] and uh we reduced the features because [13:18] uh because of the limitation of the [13:20] cubits we have and after that like we [13:23] encoded the feature using angular [13:25] encoding into the quantum states and [13:27] then we trained the model and [13:30] uh we did the comprehensive assessment [13:31] in this uh uh work. So these are the [13:35] metrics of the model like what we got [13:36] like accuracy was around 95%. [13:39] Which if you will see it looks promising [13:42] like but we'll not say that it is very [13:46] promising how it should be because uh at [13:50] times like when you're working with [13:52] quantum models if we don't get the [13:53] circuit optimized or built right the [13:55] chances are very high that these [13:56] accuracies can take huge step also at [13:59] times. So finding the right optimized [14:02] circuit is very important and what kind [14:04] of optimizers we are using over there. [14:07] Uh so when we analyze the result so we [14:09] analyze with different data set size as [14:11] well. So when we use 75% data set we got [14:14] 95 uh 94.8% accuracy. When we stopped [14:18] the when we reduce the data to 50% we [14:21] got the accuracy of 93.5%. [14:24] So you can see quantum computer doesn't [14:26] need huge data here right with the even [14:28] with the small amount of data also we [14:30] can train the generic model which is [14:32] still predicting the results with higher [14:35] accuracy in comparison to like having [14:37] huge data set. So this shows the [14:40] generalization ability with like minimal [14:42] training data. And then if we talk about [14:45] computational efficiency we have uh like [14:48] we were able to train the model quantum [14:49] model with 25 epox with better results [14:52] what we were getting with 40 epox of [14:54] classical training. Training time was [14:57] also reduced while working with it and [14:59] memory usage energy consumption [15:01] inferencing speed also like sort of like [15:03] we tried evaluating over there. So some [15:05] of the key findings for us were like [15:07] quantum layer is significantly boosting [15:09] the robot robustness to noisy and [15:12] highdimensional like inputs and it shows [15:15] like performance across all metrics [15:17] compared to classical approaches and we [15:20] can sort of like in future if we have [15:22] more reliable quantum computers and if [15:24] we start adapting the error correction [15:26] algorithms as well with the current [15:28] approaches and different kind of [15:29] algorithms we can expect much better [15:31] results u like where we are at the [15:35] moment. [15:37] Uh so what does the f uh future outlook [15:40] look like for the quantum machine [15:42] learning? It's it's still at a very uh [15:45] early stage. It's um I don't know any of [15:48] the solutions has been deployed in the [15:49] production. A lot of research is [15:50] happening definitely a lot of papers and [15:53] IPs you will see uh if you follow nature [15:56] stringer and the other publishers. [15:58] But at the same time we have huge [16:00] challenges. First challenge is like [16:02] error correction. So quantum computers [16:04] are prone to some sort of errors be it [16:07] because of the surrounding or because of [16:09] like how the code is getting compiled [16:11] and the other reasons. So how we can [16:13] sort of like adapt those error [16:14] correction algorithms? Uh that is uh one [16:16] thing u we are trying to like overcome. [16:20] We have a limited number of cubits. So [16:23] you keep hearing that some company has [16:26] launched some chip with hundred or [16:28] thousands of cubits but those are the [16:31] actual logical cubits which we can use [16:33] for these kind of algorithms and the [16:35] logical cubit counts are still less. So [16:38] which defines like how well we can [16:39] create those circuit and how we can map [16:41] the data on the quantum state and other [16:45] advancement will be like a couple of [16:47] companies are already working on like [16:49] combining LLM natural language [16:51] processing with quantum and build [16:53] training the model on top of that. So if [16:55] you can think of like how we can combine [16:56] quantum with chai integration so that [16:58] some of the complex problems can be [17:00] faster uh can be solved faster in an [17:02] optimized way that will be like uh one [17:05] added advantage over there. [17:08] uh so these are some of the algorithms [17:11] like which exist on the quantum side as [17:14] well. So we have QN as we have KN&N in [17:16] the classical side we have quantum K [17:18] nearest neighbor we have QSVM which is [17:20] the quantum version of support vector [17:22] machine algorithms and many more [17:24] algorithms you can see in the list here [17:26] at the same time with these quantum [17:28] machine learning algorithms there are [17:30] challenges associated as well. So how we [17:32] do how do we load the training data to [17:34] the circuit where people get stuck at [17:36] finding the right kind of encoding for [17:38] your data. If you don't understand the [17:39] quantum principles behind those, you [17:42] might not end up uh using the right [17:45] encoding techniques or the optimizers [17:47] and you'll not get the good results [17:48] then. So how we can avoid baron pal [17:51] platue is speed up like in theory uh [17:55] researchers like me promise a lot of [17:56] good results but when it comes to the [17:59] practicality at times like we fail to [18:00] prove in comparison to the how how [18:04] accurate results we get from the deep [18:05] learning models itself. So still a lot [18:08] of things are theoretical. We are trying [18:09] to get there. We are still not there but [18:11] sort of like the foundation is getting [18:13] laid down and people are building PC's [18:14] and experimenting around to see what can [18:17] be done and privacy security ethics like [18:21] those are going to be there whatever we [18:23] have in the AI space. [18:25] So um the talk title was not just a [18:27] metaphor to um but it is a call for [18:31] interdisciplinary collaboration. So and [18:33] it's not just for healthcare. This talk [18:35] was specific for for healthcare. So [18:38] quantum physicists, clinicians, data [18:40] scientists can come together and try to [18:42] understand like what advantage we can [18:43] gain for some of the use cases with the [18:46] help of quantum. But definitely it has [18:48] the applications in other domains as [18:50] well. So be it aerospace, e-commerce, [18:52] energy wherever optimization problem [18:54] lies. Kevin just take uh gave the talk [18:56] before me on the optimization problem. [18:58] So such kind of problems uh postquantum [19:01] cryptographic problems how we can like [19:02] stay quantum proof from those attacks [19:06] telecommunication u material discovery [19:09] there are huge use cases and companies [19:10] are working towards that [19:13] and um these are some of the leading [19:15] companies TW is also there Google INQ [19:18] IBM quantum and many other like have uh [19:22] this is not just exhaustive list these [19:23] are like some of the key players but [19:25] there are many other companies who have [19:26] been leading the uh [19:28] this work. [19:30] So, um that's it from my side. [19:33] This is the QR for the presentation. You [19:35] can scan it up and I'm open to the [19:38] questions now if we have time.