Papers
Topics
Authors
Recent
Search
2000 character limit reached

Weakly Supervised Lesion Localization With Probabilistic-CAM Pooling

Published 29 May 2020 in cs.CV | (2005.14480v1)

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.

Authors (4)
Citations (60)

Summary

  • The paper introduces PCAM pooling, which refines traditional CAM by incorporating a probabilistic approach to enable effective lesion localization with weak supervision.
  • It demonstrates that a ResNet-34 model on the ChestX-ray14 dataset outperforms existing methods in both classification and localization of thoracic diseases.
  • The method enhances the delineation of lesion boundaries while maintaining computational efficiency without adding extra trainable parameters.

An Overview of "Weakly Supervised Lesion Localization With Probabilistic-CAM Pooling"

In the paper "Weakly Supervised Lesion Localization With Probabilistic-CAM Pooling," the authors address a significant challenge in the domain of medical imaging—namely, the localization of thoracic diseases in chest X-ray images using weak supervision. The authors propose a novel global pooling operation called Probabilistic Class Activation Map (PCAM) pooling. This method facilitates lesion localization with only image-level supervision, an essential feature considering the difficulty and time consumption associated with the manual annotation of bounding boxes required for more traditional, strongly-supervised approaches.

Methodology

At the core of this work is the enhancement of the traditional Class Activation Map (CAM) framework, which has been widely used to localize lesions on chest X-rays. The authors argue that CAM is mostly applied as a post-processing technique and propose leveraging it during training through the development of PCAM pooling.

PCAM pooling enhances CAM by introducing a probabilistic interpretation during training. It employs a fully convolutional network backbone to generate feature maps from input chest X-ray images. Each feature embedding is processed to produce a class activation score, which is then bound by a sigmoid function to output lesion probability maps. These maps are normalized following the multiple-instance learning (MIL) framework to calculate attention weights for pooling via a weighted average, an approach that does not add extra trainable parameters, thereby preserving simplicity and computational efficiency.

During inference, PCAM pooling utilizes these probability maps for lesion localization, conducting a probability thresholding to delineate lesion boundaries with bounding boxes. This strategy offers improved visualization in comparison to the standard CAM, particularly in terms of boundary sharpness and clarity around lesion regions.

Results and Evaluation

The experimental evaluation is performed on the public ChestX-ray14 dataset, comprising over 100,000 images with image-level labels and only a fraction annotated with bounding boxes. By implementing PCAM pooling, the researchers show that a ResNet-34 model outperforms other state-of-the-art methods on both classification and localization tasks. The new approach improves on baseline classification and localization accuracy measures, particularly excelling in diseases like Cardiomegaly and Pneumothorax, where it achieves considerable improvements.

Of note, the generated probability maps with PCAM pooling demonstrate superior delineation of lesion boundaries when visually inspected against class activation maps produced by techniques such as LSE pooling.

Implications and Future Work

The introduction of PCAM pooling has potential implications for the further development of weakly supervised learning in medical imaging. Its ability to perform effective lesion localization with just image-level labels could significantly alleviate the burden of extensive data annotation. This capability is paramount in scaling AI-driven medical tools to larger datasets while maintaining practical feasibility.

The paper states an area for future improvement: reducing the tendency of PCAM pooling to slightly enlarge regions of interest, which can lead to higher false positives, especially for small lesions. Addressing this will likely involve the refinement of the probability maps' precision in distinguishing lesion boundaries more closely aligned with actual clinical findings.

In conclusion, PCAM pooling is a notable contribution to the field of thoracic disease diagnosis in chest X-ray imaging, demonstrating potential for wide adoption and integration into clinical diagnostic workflows. This work may usher in further exploration of probabilistic interpretations in pooling operations across other neural network architectures and applications in AI-driven diagnostics.

Paper to Video (Beta)

No one has generated a video about this paper yet.

Whiteboard

No one has generated a whiteboard explanation for this paper yet.

Open Problems

We haven't generated a list of open problems mentioned in this paper yet.

Collections

Sign up for free to add this paper to one or more collections.