[Paper Review] Weakly Supervised Medical Diagnosis and Localization from Multiple Resolutions

Yao Li, Jordan Prosky+3 more|arXiv (Cornell University)|Mar 21, 2018

COVID-19 diagnosis using AIMedicine32 references97 citations

TL;DR

The paper presents a multi-resolution weakly supervised framework that learns to diagnose and localize chest X-ray abnormalities using only image-level labels, with a novel LSE-LBA pooling for adaptable saliency maps.

ABSTRACT

Diagnostic imaging often requires the simultaneous identification of a multitude of findings of varied size and appearance. Beyond global indication of said findings, the prediction and display of localization information improves trust in and understanding of results when augmenting clinical workflow. Medical training data rarely includes more than global image-level labels as segmentations are time-consuming and expensive to collect. We introduce an approach to managing these practical constraints by applying a novel architecture which learns at multiple resolutions while generating saliency maps with weak supervision. Further, we parameterize the Log-Sum-Exp pooling function with a learnable lower-bounded adaptation (LSE-LBA) to build in a sharpness prior and better handle localizing abnormalities of different sizes using only image-level labels. Applying this approach to interpreting chest x-rays, we set the state of the art on 9 abnormalities in the NIH's CXR14 dataset while generating saliency maps with the highest resolution to date.

Motivation & Objective

• Motivate using multi-resolution analysis to improve localization of varied-size chest X-ray findings.
• Develop a weakly supervised framework that produces high-resolution saliency maps without ROI annotations.
• Introduce a learnable Log-Sum-Exp pooling with lower-bounded adaptation (LSE-LBA) to handle varying lesion sizes.
• Demonstrate state-of-the-art classification and localization on NIH Chest X-ray (CXR14) dataset.

Proposed method

• Propose a multi-resolution architecture that fuses features from high and low resolutions using dense connections. - Reduce spatial resolution with ResNet blocks and preserve resolution with DenseNet-style connections. - Iteratively upsample and concatenate features across resolutions to form coarse-to-fine representations for localization.
• Use a weakly supervised MIL framework where image-level labels supervise per-instance predictions across a 2D saliency map.
• Introduce LSE-LBA pooling: p = (1/(r0+exp(beta))) * log( (1/wh) * sum exp[(r0+exp(beta)) * S_ij] ), with r = r0 + exp(beta) providing a learnable, lower-bounded sharpness prior.
• Apply sigmoid(WI_n(x)) to obtain per-instance class probabilities, then pool via LSE-LBA to yield image-level predictions and train with multi-label cross-entropy.
• Train from scratch on NIH Chest X-ray dataset with standard data augmentation and Adam optimization; evaluate using AUC per abnormality and continuous Dice for localization.

Experimental results

Research questions

• RQ1 Can multi-resolution, weakly supervised learning produce accurate pathology localization from only image-level labels in chest X-rays?
• RQ2 Does the LSE-LBA pooling provide robust, high-resolution saliency maps across abnormalities of different sizes?
• RQ3 How does multi-resolution fusion affect classification performance (AUC) and localization accuracy (Dice) on NIH CXR14?
• RQ4 What is the impact of the sharpness prior parameter r0 on localization and classification across various pathologies?

Key findings

• Achieves state-of-the-art or competitive AUC on 9 of 14 abnormalities on NIH Chest X-ray test set without using localization labels or pretraining on ImageNet for most cases.
• Produces high-resolution probabilistic saliency maps showing improved localization for both focal and diffuse abnormalities as r0 increases.
• Localization performance (Dice) generally improves with a moderate sharpness prior (r0 around 5) and may degrade for very large r0 on diffuse abnormalities.
• Classification performance is robust to r0, while localization is more sensitive to r0, with best results for a balance between sharpness and coverage.
• Outperforms prior weakly supervised methods on several abnormalities and approaches or matches state-of-the-art on others without extra labeled data.