ABSTRACT Skin Cancer is one of the fastest‐growing cancers, and it requires early diagnosis for effective treatment. The diagnosis of skin cancer depends heavily on image segmentation, as traditional models frequently have challenges capturing the intricacy and diversity of lesion features. This research proposes a new approach to improve segmentation accuracy by combining evolved curriculum learning, multimodel training, and ConvLSTM‐based refinement. The lesion complexity (size, contrast, texture, and borders) is used to stratify the dataset into easy, moderate, and tough categories. After that, specialized models are trained: UNet for easy lesions, UNet++ for moderate lesions, and Attention UNet for tough lesions. The same image is processed by each model during inference, and the outputs are weighted by confidence masks that represent the dependability of the models. These outputs are further integrated by a ConvLSTM refinement module, which uses temporal and spatial connection to provide precise and cohesive segmentation masks. The method outperforms existing methods in validation on the ISIC 2017, ISIC 2018, and PH2 datasets. The curriculum‐based method ensures directed learning and keeps simpler instances from dominating the training process. This research paves up the possibility for sophisticated automated diagnostic tools in clinical practice by showing how curriculum learning combined with multimodel refinement may increase the robustness of medical image segmentation.
Verma et al. (Tue,) studied this question.