Impact of Fuzziness for Skin Lesion Classification with Transformer-Based Model

Skin lesion is one of the most commonly encountered illnesses that need to be detected and treated at an early stage. Numerous Convolutional Neural Network (CNN) classifiers were utilized to identify skin conditions using visual features. When it comes to computational effectiveness and accuracy, the recently released Vision Transformers (ViT) have been shown to outperform CNNs. Since the ViTs perform well in classifying images, we use ViT models in our approach. This research proposes a fuzziness-based transformer model to classify images. The divide-and-conquer technique is employed to examine the effectiveness of fuzzy logic and classifying images more accurately. To increase the effectiveness of image classification, the testing dataset has been divided into three groups, including low, medium, and high-fuzzy instances, based on the degree of fuzziness for each sample. For better classification of images, we propose a straightforward mechanism that consists of a fuzzy divide-and-conquer strategy with ViT. Experiments are carried out to compare the model’s performance with and without using this proposed mechanism. We noticed that retraining classifiers with low fuzzy samples can enhance image classifier performance. The classification performance on the PAD-UFES-20 dataset using this strategy revealed substantial improvements in the classifier’s performance.