Lightweight Deep Models for Dermatological Disease Detection: A Study on Instance Selection and Channel Optimization

The identification of dermatological disease is an important problem in the worldwide according with different studies. Several works in literature use the datasets of different repositories without applying a study of the data behavior, especially in medical images domain. In this work, we propose a methodology to preprocess dermaMNIST dataset in order to improve its quality for the classification stage, where we use lightweight convolutional neural networks. Experimental results show that our method significantly reduces the number of training instances required while maintaining a classification performance comparable to that of more complex architectures such as ResNet-50 and ResNet-101 (our model obtains 71.52\%, while ResNet-50 and ResN2t-101 reach 71.72\% and 70.07\%, respectively). According to the Kruskal–Wallis hypothesis test, the performance differences between our lightweight models and the ResNet baselines are not statistically significant (p-value=0.91), highlighting the effectiveness and efficiency of our proposed pipeline.