[Paper Review] Weakly Supervised Lesion Localization With Probabilistic-CAM Pooling

Wenwu Ye, Jin Yao+2 more|arXiv (Cornell University)|May 29, 2020

COVID-19 diagnosis using AIMedicine1 references47 citations

TL;DR

This paper introduces Probabilistic-CAM (PCAM) pooling, a weakly supervised global pooling method that uses CAM localization during training to improve lesion localization on chest X-rays without bounding-box annotations, achieving superior classification and localization compared to baselines on ChestX-ray14.

ABSTRACT

Localizing thoracic diseases on chest X-ray plays a critical role in clinical practices such as diagnosis and treatment planning. However, current deep learning based approaches often require strong supervision, e.g. annotated bounding boxes, for training such systems, which is infeasible to harvest in large-scale. We present Probabilistic Class Activation Map (PCAM) pooling, a novel global pooling operation for lesion localization with only image-level supervision. PCAM pooling explicitly leverages the excellent localization ability of CAM during training in a probabilistic fashion. Experiments on the ChestX-ray14 dataset show a ResNet-34 model trained with PCAM pooling outperforms state-of-the-art baselines on both the classification task and the localization task. Visual examination on the probability maps generated by PCAM pooling shows clear and sharp boundaries around lesion regions compared to the localization heatmaps generated by CAM. PCAM pooling is open sourced at https://github.com/jfhealthcare/Chexpert.

Motivation & Objective

• Motivate accurate localization of thoracic lesions using only image-level labels to enable scalable training.
• Leverage CAM-based localization during training via a probabilistic pooling scheme to guide feature aggregation.
• Achieve improved classification and localization performance on a large chest X-ray dataset without extra annotation effort.

Proposed method

• Backbone network processes chest X-ray images to produce feature maps.
• Compute class activation scores per embedding using a 1x1 convolution; bound with sigmoid to get per-embedding disease probability.
• Normalize these probabilities to obtain attention weights under a MIL framework for weighted global pooling.
• Pool the feature map with these attention weights to produce an embedding that goes through the same classification layer.
• During inference, use the sigmoid-based probability maps as localization heatmaps for region/probability thresholding to obtain regions and bounding boxes.

Experimental results

Research questions

• RQ1 Can PCAM pooling improve lesion localization under image-level supervision compared to existing weakly supervised methods?
• RQ2 Does integrating CAM-based localization into the training objective enhance both classification and localization performance on chest X-ray datasets?
• RQ3 What is the qualitative impact of PCAM probability maps on localization boundaries relative to standard CAM heatmaps?

Key findings

• PCAM pooling improves both classification AUC and localization metrics over several baselines on ChestX-ray14.
• The probability maps produced by PCAM pooling exhibit clearer, sharper boundaries around lesions than traditional CAM heatmaps.
• Localization experiments show PCAM pooling achieves higher IoBB-based localization accuracy across multiple diseases, with some trade-offs in false positives.
• Qualitative visualizations demonstrate more precise lesion localization boundaries and broader region suggestions than LSE-based approaches.