A Multiscale Convolutional Neural Network Framework for Automated Segmentation and Pattern Mapping of Psoriatic Lesions

For psoriatic lesions, automatic segmentation is crucial to perform objective assessment, monitoring, and medication planning in dermatology. This study proposes a Multiscale Convolutional Neural Network (MSCNN) framework for precise segmentation of psoriatic lesions from medical images. The model was evaluated using the ISIC 2017 dataset and demonstrated robust performance in lesion localization. By analyzing the predicted masks, it has been observed that the boundaries closely match ground truth annotations. Metrics for quantitative evaluation are Dice Coefficient, Intersection over Union (IoU), used with high precision and slightly lower recall, reflecting occasional under-segmentation of fine-scale lesion details. The findings focus on the proposed MSCNN’s capability to produce reliable lesion masks, while also detecting areas for advancement in capturing irregular lesion boundaries. Future work involves integrating multimodal imaging, attention mechanisms, and larger, diverse datasets to improve segmentation accuracy and clinical applicability.