Lightweight Multi-task CNN for Otoscopic Image

Otitis media is a common middle-ear inflammatory disease that often leads to hearing loss if not diagnosed in time. However, accurate diagnosis is limited by the availability of specialists and medical imaging equipment. This study proposes a lightweight multi-task convolutional neural network (CNN) that simultaneously performs classification and lesion localization on otoscopic images, providing interpretable visual feedback for clinical use. The proposed framework consists of a shared lightweight encoder and a compact decoder that produces lesion probability maps. The model is trained on a public otoscopy dataset [1], covering a range of clinically relevant diagnostic categories. A hybrid loss combining cross-entropy and Dice objectives enables stable training without manual segmentation labels. Experimental results show that the network achieves consistent accuracy with low computational demand, producing clear lesion bounding box that assist specialists in explaining diagnostic results to patients. Because of its minimal model size and efficient inference, the proposed system can be deployed on mobile or embedded platforms, enabling rapid screening, telemedicine consultation, and future integration into smart otoscope devices for home healthcare. This work contributes to the development of practical and ubiquitous intelligent diagnostic tools for otolaryngology.