Does multi-task supervision teach a classifier where to look? A quantitative attention-localization study in brain tumor MRI

Background: Multi-task learning is often assumed to improve clinically relevant attention, but this is largely unproven. Methodology: An EfficientNet-B4 multi-task model classified 30 pathology classes and predicted lesion coordinates on brain MRI; Grad-CAM++ centroids were compared to radiologist annotations on a 1,128-image validation set. The study quantitatively tests whether a lesion-localization head aligns diagnostic attention with true tumor centers across pathology classes.