Beyond Detection: Comparative Explainability Study on Trypanosoma cruzi Using CAMs and DETR Attention

Aqsa Yousaf, Paul Agbaje, Megan Coffee, Habeeb Olufowobi

Confirmation: I have read and agree with the IEEE BHI 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: Explainable AI, Chagas disease, Trypanosoma cruzi, Class Activation Mapping, Medical image analysis, Blood parasites, Micorscopy

Abstract: Chagas disease, caused by Trypanosoma cruzi, demands accurate and interpretable detection methods to support clinical decision-making. While deep learning models such as YOLOv8 and DINO-DETR perform well on microscopy images, their lack of interpretability hinders clinical adoption. We present the first comparative Explainability study of CNN- and transformer-based object detectors for Trypanosoma cruzi detection. For YOLOv8, we benchmark ten Class Activation Mapping explainable AI (CAM-XAI) methods across multiple internal layers, evaluating interpretability using Intersection-over-Union (IoU) and Energy-Based Pointing Game (EBPG). For DINO-DETR, we introduce a query-specific attention visualization method that maps decoder attention of a query to image space without backpropagation. Our results reveal complementary behaviors: CAMs highlight broad parasite regions, while DETR attention targets fine-grained, discriminative features. We further demonstrate that existing localization metrics are inadequate for shared heatmaps in multi-object settings, underscoring the need for new localization evaluation metrics in medical explainability.

Track: 4. Clinical Informatics

Registration Id: 2KNQRF7YD2C

Submission Number: 126