Reptile Optimization-Based Framework for Precise Lesion Segmentation in Skin Images

Abstract

Precise lesion segmentation in skin images is critical for the early detection and diagnosis of skin cancers, particularly melanoma. Existing lesion segmentation methods often struggle with the variability in lesion shapes, sizes, and the low contrast between lesions and their surrounding skin. Furthermore, they typically require large annotated datasets and perform poorly when applied to unseen lesion types or domains. To address these challenges, we propose the Reptile Optimized Lesion Segmentation Network (ROLSe-Net), which integrates the Reptile meta-learning algorithm with a U-Net backbone. This framework optimizes model initialization across multiple lesion segmentation tasks, enabling rapid fine-tuning on new data with minimal supervision. Additionally, the framework incorporates data augmentation and post-processing steps to enhance boundary precision. ROLSe-Net is applied to automate the segmentation of melanoma lesions in dermoscopic images, aiming to support dermatologists in making informed clinical decisions. Experimental results on benchmark datasets demonstrate that ROLSe-Net achieves superior performance compared to traditional segmentation methods, especially in low-data and crossdomain scenarios. The proposed method significantly improves segmentation accuracy, boundary delineation, and adaptability to new lesion types, showcasing its robustness and practical relevance.