Squeeze-and-Excitation Channel Attention in a Compact CNN for 24-Class Clinical Skin Disease Classification

Access to dermatologists is limited in many low-resource regions, which creates a need for automated triage tools that are accurate yet light enough for modest hardware. This paper presents a compact convolutional neural network (CNN), trained from scratch, that places a Squeeze-and-Excitation (SE) channel-attention block in each of its four convolutional stages. From a public collection of clinical skin-disease photographs, a class-balanced dataset of 68,568 images in 24 classes (23 disease categories and healthy skin) was built by randomly undersampling every category to 2,857 images. On a held-out split of 13,728 images, the SE-CNN (5.01 M parameters) reached 88.55% top-1 accuracy and a macro F1 of 0.888, with a micro-averaged ROC AUC of 1.00. In a controlled ablation on a reduced subset, adding SE blocks to an otherwise identical network raised validation accuracy from 64.2% to 73.1% and test macro F1 from 0.49 to 0.59 while adding less than 1% to the parameter count. On the same subset, the SE-CNN also exceeded MobileNetV2 and DenseNet121 trained under the same protocol. The paper closes with an explicit account of the evaluation limits of this study, including image-level splitting and the reuse of the held-out split for early stopping, and the experiments needed before clinical use.