Rare-Subtype Performance in Intracranial Hemorrhage Classification Is Protocol-Dependent: trained model weights

Trained model weights for a systematic re-evaluation of rare-subtype performance in intracranial hemorrhage classification from CT. 46 checkpoints:• Stage-1 slice classifiers — four architecture families (ConvNeXt-tiny, ResNet50, EfficientNet-B4, DenseNet-161) under a logit-adjusted loss• Stage-1 classifiers under focal loss and class-weighted cross-entropy• Stage-2 conditional and joint study-level heads• The re-implemented base method, both training passes All were trained on the patient-disjoint partition released with the code archive (hash f72016749123e8ce). Weights only; no patient data.