Durar Alqahtani

Artificial Intelligence Specialization · Cohort 2026

Explainable Knowledge Distillation for Bone Fracture Classification Using X-ray Images

Bone fracture classification from X-ray images is an important application of deep learning for supporting clinical diagnosis. However, high-capacity models are often computationallyexpensiveanddifficulttointerpret,limitingtheirdeploymentinresource- constrained healthcare environments. This study investigates an Explainable Knowledge Distillation(EKD)frameworkthatcombinesprediction-basedknowledgedistillationwith class-discriminative attention transfer to develop a lightweight and interpretable fracture classifier. Using the FracAtlas dataset, EfficientNet-B6 was selected and optimized as the teacher model, while MobileNetV2 was selected as the lightweight student. The EKD student was evaluatedunder three training settings: No Augmentation, RandAugment, and AugMix, and compared with the corresponding MobileNetV2 baselines. EKD improvedvalidationaccuracyacrossallthreesettings,withAugMixachievingthehighest validationaccuracyof0.8742. Qualitativecomparisonsofteacherandstudentattention maps showed partial spatial correspondence,althoughthe student maps were generally broaderandmorediffuse. Overall,theresultsdemonstratethepotentialofexplainability- guided knowledge distillation for improving lightweight fracture classification while encouraging similarity between teacher andstudent attention patterns.

Read the final report (PDF)