A Comparative Study of Lightweight Deep Learning Models for Skin Lesion Classification: Performance, Computational Efficiency and Deployment Considerations

Deep learning is now the standard way to classify skin lesions from images, because it learns features straight from thermoscopic images and nobody has to design them by hand. The trouble is size. Large convolutional neural networks (CNNs) are hard to run on smartphones, embedded medical devices and other edge hardware, where memory and energy are scarce and latency matters. This review looks at five lightweight CNNs for skin-lesion classification (SqueezeNet, MobileNetV2, MobileNetV3, ShuffleNetV2 and EfficientNet-B0) and asks how their predictive performance relates to their computational cost. Working from the published literature, we explain what makes each model small and put the reported classification and efficiency figures side by side in tables. We also cover the HAM10000 dataset and its heavy class imbalance, data-splitting protocols, transfer learning, augmentation, parameter count, FLOPs/MAdds, model size and latency measured on real devices. The studies we reviewed show that lightweight networks can reach useful accuracy, but their numbers come from different datasets and protocols, so they are not results from one common benchmark. One recent study reports 97.05% accuracy next to an 89.91% macro-F1, so a high headline accuracy can sit on top of weak results for the rare classes. That is why we give macro-averaged metrics and predictive uncertainty so much weight. To deal with this, we define a reproducible framework for a future controlled benchmark. Every candidate model would be trained and evaluated under the same data, preprocessing, training and hardware conditions, and the evaluation metrics are defined mathematically. This is a Major Project-I proposal, so it reports no new experimental or clinical results, and every number in it comes from a cited study. What it does offer is a fair way to measure the accuracy-efficiency trade-off and a list of the conditions a deployment-oriented comparison would need.