Mask-aware knowledge distillation for low-latency pneumothorax classification in chest radiographs

Accurate pneumothorax classification requires strong discrimination with low deployment cost. We developed a mask-aware knowledge distillation framework in which a mask-guided ResNet-50 teacher transferred dynamically annealed logits and lesion-weighted multilevel attention to a MobileNetV3-Small student, while masks and teacher-side components were removed at inference. A fixed stratified 20% SIIM-ACR hold-out set was reserved for testing; the remainder underwent five-fold cross-validation, with PTX-498 positives used only for training augmentation. The Mask-aware KD student achieved a mean validation area under the receiver operating characteristic curve (AUC) of [Formula: see text], and the five-model student ensemble achieved a fixed-test AUC of 0.946 (95% confidence interval, 0.935-0.956) versus 0.950 for the teacher ensemble. Validation-selected 90% and 95% sensitivity targets yielded test sensitivities of 0.929 and 0.973. Uniform-attention KD achieved a lower AUC of 0.936. The student ensemble required 5.00 million parameters, 1.49 billion multiply-accumulate operations (MACs), and 51.00-ms single-threaded CPU latency at [Formula: see text]. Training-time lesion guidance enabled numerically similar discrimination with substantially lower inference cost. No independent external validation was performed, and generalizability across institutions, acquisition protocols, portable radiography settings, and patient populations remains uncertain; verified patient-level evaluation remains necessary before clinical translation.