FM-net: focal modulation-based network for accurate skin lesion segmentation

Abstract Background: Accurate segmentation of skin lesions in dermoscopic images is essential for the detection of skin cancers, particularly melanoma, as precise boundary identification enables timely and effective diagnosis. Cancerous skin lesions, particularly malignant melanoma, contribute significantly to the mortality rate, underscoring the need for early detection and precise diagnostic techniques. However, achieving this precision poses challenges owing to indistinct lesion borders, asymmetrical shapes, and common obstructions such as hair, markings, and occlusions. Method: This study addresses these challenges by introducing an end-to-end trainable network incorporating focal modulation to enhance the feature refinement for pixel-level classification. Focal modulation effectively extracts fine-grained, multi-scale features while preserving contextual information, thereby improving the performance of lesion segmentation. The decoder part of the proposed method utilizes transposed convolution for up-sampling, which preserves the spatial detail necessary for high-resolution segmentation. Results and Conclusion: The proposed FM-Net was validated on several benchmark datasets and demonstrated state-of-the-art (SOTA) performance in skin lesion segmentation. The FM-Net’s ability to provide SOTA performance without data augmentation is one of its significant strengths. Performance on the International Skin Imaging Collaboration (ISIC) datasets demonstrated FM-Net’s adaptability, with Jaccard index scores of 89.60% on ISIC 2016, 82.34% on ISIC 2017, and 87.71% on ISIC 2018. The comprehensive performance of FM-Net highlights its ability to accurately segment lesions and its potential for assisting in timely diagnosis. The code used to reproduce the results is available on GitHub .