Weakly-supervised explainable infection severity classification from chest CT scans

TL;DR

This paper introduces a weakly-supervised method for classifying infection severity in chest CT scans, offering explainable results to aid clinicians in treating respiratory diseases.

Contribution

A novel weakly-supervised classification approach for infection severity with explainable results, avoiding expensive segmentation annotations.

Findings

The proposed method achieves state-of-the-art performance on multi-centre SARS-CoV-2 datasets.

It shows significant performance gains on cross-site train/test splits.

The approach provides explainable infection coverage through fused low- and high-level features.

Abstract

Novel respiratory diseases can have a devastating impact on healthcare systems, which underlines the importance of developing methods to improve the prevention, diagnosis, and prognosis of such diseases. Developing computer-aided diagnosis tools that determine infection severity can aid healthcare professionals in deciding treatment strategies and preventing cross-infection. In such manner, lung infection severity classification from chest CTs is crucial for deciding treatment plans and interventions needed to block illness progression in individual patients and reduce cross infection. However, current techniques face performance, generalizability, and explainability challenges for automated infection severity classification methods due to the high spatial complexity of 3D volumes. Significant efforts have been focused on segmentation approaches to quantify lung infection and assess infection severity, but such segmentation methods require expensive data annotation and clinical expertise. In this work, we propose a weakly-supervised classification approach to distinguish between different levels of infection, while providing clinicians with explainable results. To mimic clinical practice, the different stages in our approach focus on low-level infection patterns as well as high-level infection coverage in lung CT scans. We then fuse the high-level features with the positionally encoded low-level features to provide volume-level infection classification. Testing on the SARS-CoV-2 (COVID-19) multi-centre multi-region datasets, our approach shows promising performance gains compared to existing state-of-the-art methods, where we achieve state-of-the-art severity classification performance. Furthermore, we show significant performance gains on cross-site train/test splits. Finally, we quantitatively and qualitatively demonstrate the explainability of our weakly-supervised approach, where we can achieve substantial infection coverage.

Figures