U-NetSkinLesionNet++: A Hybrid Custom CNN– ViT Model with CycleGAN-Augmented K-means Segmented Data for Enhanced Skin Cancer Detection and Classification

Abstract

As skin cancer remains the most frequent type of cancer worldwide, patient outcomes need accurate diagnostic techniques. In the following paper, an in-depth comparison of advanced segmentations with trusted deep learning models for the classification of skin cancer images is presented. For the preparation of an original segmented dataset, the original dataset was initially preprocessed and then segmented into K-means clustering. It was subjected to the application of deep learning architectures, including existing pre-trained models such as DenseNet201, ResNet50, ConvNeXt_Base, EfficientNetV2-L, InceptionV3, Xception, and Swin Transformer-B, and an innovatively developed hybrid CNN– ViT model, in the subsequent steps. CycleGAN was employed for efficient augmentation of the dataset in order to solve issues related to class imbalance, which is typical for medical data. The same deep learning models, including DenseNet201, ResNet50, ConvNeXt_Base, Swin Transformer-B, and the proposed hybrid CNN–ViT model, were employed for the development of U-Net++ in order to construct the complex segmentation method. With 90% accuracy, the proposed hybrid CNN–ViT model performed best among them. Various quantitative assessments, including confusion matrix, loss curves, ROC curves, and accuracy, indicate the potential for the proposed model if combined with CycleGAN-based augmentation. Comparative results indicated that while U-Net++ segmented dataset models achieved up to 88.2% accuracy, 89% accuracy was achieved in Kmeans clustering segmentation. Valuable insights into the relative superiority of K-means compared with U-Net++ segmentation methods are provided through the comparison work, in addition to providing important tips for the selection of the best preprocessing methods in clinical scenarios. The conclusions in the paper have enormous possibilities for enhancing computerized skin cancer detection, and it has the prospect of contributing towards enhanced survival rates in patients, as well as early detection rates.