SE-Fundusnet: A Channel Attention-Based Deep Learning Architecture for Papilledema Detection and Classification

Papilledema is characterized by enlargement of the optic disc resulting from increased intracranial ‎pressure, and represents a critical clinical finding that demands immediate medical attention. A new deep ‎learning method for automatically identifying papilledema from fundus images is presented in this work. ‎A Hybrid Centric convolutional neural network (HCCNN) architecture is introduced that incorporates ‎squeeze-and-excite (SE) channel attention mechanisms and parallel feature extraction to effectively ‎capture diagnostically relevant features. The model employs different kernel sizes across parallel ‎processing pathways to simultaneously extract fine-grained details and broader contextual features. ‎Experimental results demonstrate excellent performance, with an overall accuracy of 96.98%, sensitivity ‎of 95.48% and specificity of 97.74%. Ablation studies confirm the effectiveness of the SE blocks, which ‎improve detection accuracy by 4.3%, particularly for subtle presentations. This work represents a ‎significant step toward developing reliable computer-aided diagnostic tools for papilledema, potentially ‎enhancing early detection and monitoring of this vision-threatening condition‎.