Iterative Explainability for Weakly Supervised Segmentation in Medical PE Detection

TL;DR

This work tackles the scarcity of fine-grained PE annotations by introducing iExplain, an iterative, weakly supervised framework that converts image-level PE labels into pixel-precise segmentation via explainability-driven pseudo-labels. The pipeline combines a slice-level classifier, iterative Integrated Gradients-based pseudo-label generation with hysteresis thresholding, and a DynUNET-based segmentation model, achieving strong localization while avoiding auxiliary models. On the RSPECT-augmented dataset, the method yields competitive results with strongly supervised approaches (e.g., F1 ≈ 71.6% in baseline, rising to 75.5% with limited human annotation for finetuning), demonstrating data-efficient learning for PE detection and localization. The approach offers a practical path to high-performance weakly supervised segmentation with broad potential for application in other medical imaging tasks and beyond.

Abstract

Pulmonary Embolism (PE) are a leading cause of cardiovascular death. Computed tomographic pulmonary angiography (CTPA) is the gold standard for PE diagnosis, with growing interest in AI-based diagnostic assistance. However, these algorithms are limited by scarce fine-grained annotations of thromboembolic burden. We address this challenge with iExplain, a weakly supervised learning algorithm that transforms coarse image-level annotations into detailed pixel-level PE masks through iterative model explainability. Our approach generates soft segmentation maps used to mask detected regions, enabling the process to repeat and discover additional embolisms that would be missed in a single pass. This iterative refinement effectively captures complete PE regions and detects multiple distinct embolisms. Models trained on these automatically generated annotations achieve excellent PE detection performance, with significant improvements at each iteration. We demonstrate iExplain's effectiveness on the RSPECT augmented dataset, achieving results comparable to strongly supervised methods while outperforming existing weakly supervised methods.

Figures (5)

• Figure 1: iExplain, our weakly supervised algorithm for PE detection. 3D whole image mini-volumes (7 slices) are classified; positive cases generate PE probability heatmaps via Integrated Gradients. These are converted to clusters using Hysteresis thresholding, clusters are masked out in the input, and the process iterates until negative classification. Complete study predictions are obtained via sliding window and union of all detected clusters.
• Figure 2: Iterative refinement examples showing iExplain's PE discovery progression: A) Original study B) Ground Truth C) First iteration D) Final iteration. Our method detects both proximal PEs (rows 1-2) and PEs in the opposite lung (rows 3-4), demonstrating effective discovery of multiple embolisms through progressive masking.
• Figure 3: Mismatch example between red bounding box annotations and blue PE segmentations. While bounding boxes work well for object detection evaluation, they're inadequate for segmentation evaluation. Therefore, we use F1 score for bounding box matching rather than segmentation Dice.
• Figure 4: Left: F1 performance evolution through iterations of iExplain. Sensitivity increases while PPV decreases as new PEs are discovered, reaching optimal F1 after 6 iterations before slight decline as metrics diverge. Right: Distribution of required refinement iterations in iExplain. Most studies need few iterations ( 15% need none), with decreasing requirements as PE detection completes. The 10-iteration limit creates a tail peak of cases that would continue refining despite diminishing performance.
• Figure 5: Qualitative Error Examples. Common PE detection false positives include: A) Reduced contrast in a vein adjacent to an artery is mistaken for a PE. B) High contrast in veins due to poor contrast timing. C) Image Left: Shadows due to imaging artifact, Image Right: vein mistaken for an artery. D) Vein mistaken for an artery.