Semantic segmentation of skin lesions using deep CNNs with artifact removal and multiclass detection
Skin cancer is the uncontrolled growth of abnormal skin cells and can affect anyone. The diagnosis typically involves clinical screening, image and dermoscopic analysis, followed by biopsy and histopathological examination. Automated skin lesion classification remains challenging due to varying imag...
| Autores: | , , |
|---|---|
| Tipo de recurso: | artículo |
| Fecha de publicación: | 2025 |
| País: | España |
| Institución: | Universitat Ramon Llull (URL) |
| Repositorio: | DAU Arxiu Digital de la Universitat Ramon Llull |
| OAI Identifier: | oai:dau.url.edu:20.500.14342/5966 |
| Acceso en línea: | http://hdl.handle.net/20.500.14342/5966 https://doi.org/10.1145/3784713.3784721 |
| Access Level: | acceso abierto |
| Palabra clave: | Artificial removal Deep convolutional neural network Multiclass detection Skin lession segmentation Semantic segmentation 004 61 |
| Sumario: | Skin cancer is the uncontrolled growth of abnormal skin cells and can affect anyone. The diagnosis typically involves clinical screening, image and dermoscopic analysis, followed by biopsy and histopathological examination. Automated skin lesion classification remains challenging due to varying image quality and the presence of artifacts. Among the key steps is lesion segmentation, which is often hindered by visual artifacts such as hair, skin marks, and other noise. This study presents a clinical skin lesion segmentation method using semantic segmentation with multiclass detection. The proposed pipeline (1) artifact detection using morphological operators (2) harmonic inpainting for area restoration (3) segmentation with DeepLabv3+ with ResNet-18 backbone architecture + class weighting on 4 classes of skin, melanoma, seborrheic keratosis and nevus. Experiments were conducted on the ISIC 2017 Challenge Dataset on segmentation, which includes 2000 lesion images with superpixel masks for training, 600 image-masks pair for validation, and 150 image-masks pair for testing. Due to class imbalance, a common issue in segmentation tasks, class weighting was implemented to ensure balanced learning. The proposed method using a hybrid DeepLabV3+ model with ResNet-18 backbone architecture achieved an accuracy of 0.9143 and a weighted intersection over union (wIoU) score of 0.86307, demonstrating its effectiveness in segmenting skin lesions from clinical images. |
|---|