MalariAI: A Label-Resilient Framework for Cell Segmentation and Explainable Malaria Stage Classification

The arXiv paper introduces MalariAI, a decoupled pipeline for dense malaria blood smears. Its first stage uses annotation-agnostic watershed segmentation to isolate cells without ground-truth input, while EfficientNet-B0 with Focal Loss classifies parasite stages from cell crops. On 1600x1200 images, watershed recovered 75.95% of ground-truth cells and the end-to-end binary parasitized AP@0.5 was 29.10%. The crop classifier reached 98.36% accuracy, but rare-stage performance was lower. Grad-CAM++ supplied per-cell visual evidence, with activation energy significantly above a geometric chance baseline.