Abstract
Medical image classification often fails for two reasons. Rare but clinically important categories create class imbalance. Similarity between classes also makes some diagnoses hard to separate without a loss that focuses on fine patterns. We introduce Validation Adaptive Focal Loss (VALF), a plug and play objective that augments focal loss with per class weights that are initialized uniformly or from a user provided prior and that are adapted during training based on validation feedback. We keep the weights fixed for the initial part of training, then update them after each epoch using per class validation accuracy. We apply a small multiplicative change and then renormalize the mean weight. The loss is class weight times focal factor times cross entropy. VALF needs no architectural changes, no auxiliary network, and no multi stage training schedule. On LungHist700 at 20× and 40× across five backbones, VALF attains the top macro F1 in 8 of 10 settings and yields consistent gains in accuracy, precision, and recall. The largest macro F1 improvement is about +4.0% over the best baseline at 40×. Improvements are robust across models and magnifications, with only minor shortfalls in two 20× cases. These results indicate that simple validation driven and class aware weighting can balance sensitivity and specificity and can serve as a practical drop in for clinical pipelines.