[Paper Review] Chester: A Web Delivered Locally Computed Chest X-Ray Disease Prediction System

Joseph Cohen, Paul A. Bertin+1 more|arXiv (Cornell University)|Jan 31, 2019

COVID-19 diagnosis using AI22 references35 citations

TL;DR

A web-based prototype that runs chest X-ray disease prediction locally in the browser, with out-of-distribution detection and prediction explanation, preserving patient data privacy.

ABSTRACT

In order to bridge the gap between Deep Learning researchers and medical professionals we develop a very accessible free prototype system which can be used by medical professionals to understand the reality of Deep Learning tools for chest X-ray diagnostics. The system is designed to be a second opinion where a user can process an image to confirm or aid in their diagnosis. Code and network weights are delivered via a URL to a web browser (including cell phones) but the patient data remains on the users machine and all processing occurs locally. This paper discusses the three main components in detail: out-of-distribution detection, disease prediction, and prediction explanation. The system open source and freely available here: https://mlmed.org/tools/xray

Motivation & Objective

• Bridge the gap between deep learning researchers and medical professionals by providing an accessible, locally running chest X-ray tool.
• Demonstrate a second-opinion diagnostic aid that preserves patient privacy by processing data entirely on the client side.
• Explore three components: out-of-distribution detection, disease prediction, and prediction explanation.
• Offer an open-source model and deployment approach that can be copied to scale medical AI solutions with minimal server costs.

Proposed method

• Use DenseNet-121 for disease prediction on ChestX-ray14-derived data.
• Implement out-of-distribution detection using autoencoder-based reconstruction (ALI) and SSIM/L1/L2 metrics.
• Calibrate prediction probabilities with a piecewise linear postprocessing transformation.
• Explain predictions using gradient-based saliency maps and Class Activation Mapping.
• Deploy models in the browser via a PyTorch→ONNX→TensorFlow→TensorFlow.js pipeline.
• Provide an in-browser processing pipeline that transfers code and weights via a URL while keeping patient data on the client.

Experimental results

Research questions

• RQ1 How well can a locally run deep learning model predict chest X-ray diseases compared to existing benchmarks?
• RQ2 Can out-of-distribution detection effectively determine when not to make a prediction in a clinical context?
• RQ3 Do visual explanations (saliency maps and CAM) provide meaningful localization for radiologist interpretation?
• RQ4 Is it feasible to deploy accurate chest X-ray models entirely in the browser with acceptable performance and privacy guarantees?

Key findings

• The system demonstrates a complete, browser-based tool for chest X-ray diagnosis with OoD detection and explanation capabilities.
• Data augmentation and external validation were explored to assess generalization across datasets like PadChest.
• Out-of-distribution methods using ALI-based reconstructions and SSIM showed promise in identifying non-chest-X-ray inputs and guiding model usage.
• Gradient-based saliency maps and Class Activation Maps provide localization that highlights regions contributing to predictions.
• The deployment pipeline converts PyTorch models to ONNX and then to TensorFlow.js to operate within browsers, with validation showing close agreement to PyTorch predictions within small tolerances.