Efficient and Interpretable Otoscopic Image Classification via Distilled CNN with Adaptive Channel Attention

Accurate classification of otoscopic ear images is crucial for early diagnosis of ear pathologies such as Chronic Otitis Media, Earwax Plug, and Myringosclerosis. In this study, we propose a novel deep learning framework that employs a knowledge distillation strategy, wherein a high-capacity pre-trained teacher model (Vision Transformer or ResNet101) transfers learned representations to a lightweight student CNN model (EfficientNet-B0). The student network is further enhanced through the integration of an Adaptive Channel Attention (ACA) module, which selectively emphasizes informative features via channel-wise recalibration. The multi-scale feature distillation from the teacher improves generalization while the ACA block boosts sensitivity to clinically relevant regions. We validated our method on a publicly available otoscopic image dataset comprising four balanced classes. Our approach achieved an overall accuracy of 98.75%, with class-wise AUC values of 1.000 (CSOM), 1.000 (Earwax Plug), 0.994 (Myringosclerosis), and 0.992 (Normal), and a micro-average AUC of 0.997. Additionally, Grad-CAM analysis confirmed the model’s focus on diagnostically meaningful areas, supporting the interpretability of the predictions. These results demonstrate the effectiveness of our distillation-based ACA-enhanced architecture in otoscopic image classification, with potential to assist clinical decision-making in primary care and telemedicine applications.