This repository contains code for automatic detection of Chagas disease parasites (Trypanosoma cruzi) in blood smear images using YOLOv8, along with explainability analysis using Class Activation Maps (CAMs).
The project focuses not only on detection performance, but also on evaluating how well model explanations align with ground-truth parasite locations.
- Task: Parasite detection in blood smear microscopy images
- Model: YOLOv8
- Explainability: CAM-based methods (e.g., EigenCAM, Grad-CAM variants)
- Evaluation Metrics:
- Intersection over Union (IoU) between CAMs and ground truth
- Energy-Based Pointing Game (EBPG)
- Faithfulness Metrics (Insertion & Deletion AUC)
- KL Divergence between saliency maps and parasite bounding boxes
- Per-bounding-box and whole-image analysis
This framework helps assess whether the model is focusing on biologically relevant regions rather than spurious background cues.