Browser-Based Multi-Cancer Classification Framework Using Depthwise Separable Convolutions for Precision Diagnostics.

Area of Science:

• Artificial Intelligence
• Medical Diagnostics
• Computational Pathology

Background:

• Accurate cancer detection is crucial but hindered by deep learning's hardware and privacy demands.
• Existing AI solutions often require high-performance computing and centralized data storage.
• Data privacy concerns limit the adoption of AI in sensitive medical applications.

Purpose of the Study:

• To develop a browser-based, client-side deep learning framework for multi-cancer classification.
• To enable real-time, privacy-preserving cancer diagnosis without specialized infrastructure.
• To demonstrate the feasibility of high-accuracy AI deployment in resource-limited settings.

Main Methods:

• Fine-tuned the Xception architecture using depthwise separable convolutions on over 130,000 multi-cancer images.
• Implemented a browser-based framework using TensorFlow.js for client-side inference.
• Benchmarked performance against VGG16, ResNet50, EfficientNet-B0, and Vision Transformer models.

Main Results:

• Achieved Top-1 accuracy of 99.85% and Top-5 accuracy of 100%, outperforming all comparators.
• Demonstrated lightweight computational requirements suitable for standard web browsers.
• Grad-CAM visualizations confirmed predictions were based on relevant histopathological features, ensuring interpretability.

Conclusions:

• Presents the first fully browser-deployable, privacy-preserving deep learning framework for multi-cancer diagnosis.
• High-accuracy medical AI is achievable without significant infrastructure overhead or specialized hardware.
• Establishes a practical pathway for equitable and cost-effective global deployment of AI in cancer diagnostics.