Scholarlink

<p>Accurate segmentation of melanoma lesions in dermoscopic images has been a critical step in the early detection and diagnosis of skin cancer. In this paper, a high-performance deep learning framework has been developed based on the DeepLabv3 + architecture, integrating Atrous Spatial Pyramid Pooling (ASPP) for multi-scale feature extraction and a ResNet-101 backbone for rich contextual representation. The proposed model has been trained and validated using the ISIC-2016 dataset, with extensive experiments demonstrating superior performance in comparison to related methods. Specifically, the model has achieved an accuracy of 99.02%, a Dice Similarity Coefficient (DSC) of 93.24%, an Intersection over Union (IoU) of 87.94%, and a Boundary F1 Score (BF) of 90.63%, setting a new benchmark for lesion segmentation. To our knowledge, this is one of the first studies to systematically integrate a ResNet-101 encoder with DeepLabv3 + and conduct an extensive backbone comparison (including ResNet-50, MobileNetV2, and EfficientNet-B0) on the ISIC-2016 dataset. Additionally, performance has been validated through learning curve analysis and confusion matrices, confirming both accuracy and generalization across lesion types. The results have highlighted the model’s novelty, its ability to address critical research gaps in multi-scale and boundary-aware segmentation, and its strong potential for clinical integration, offering a reliable and scalable solution for dermatological image analysis.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A High-Performance Deep Learning Framework for Skin Lesion Segmentation Based on DeepLabv3 + and Multi-Scale Context Modeling

  • 期刊论文

  • 发表日期：

    2025年8月20日

查看全文
  • Nattavut Sriwiboon

摘要

Accurate segmentation of melanoma lesions in dermoscopic images has been a critical step in the early detection and diagnosis of skin cancer. In this paper, a high-performance deep learning framework has been developed based on the DeepLabv3 + architecture, integrating Atrous Spatial Pyramid Pooling (ASPP) for multi-scale feature extraction and a ResNet-101 backbone for rich contextual representation. The proposed model has been trained and validated using the ISIC-2016 dataset, with extensive experiments demonstrating superior performance in comparison to related methods. Specifically, the model has achieved an accuracy of 99.02%, a Dice Similarity Coefficient (DSC) of 93.24%, an Intersection over Union (IoU) of 87.94%, and a Boundary F1 Score (BF) of 90.63%, setting a new benchmark for lesion segmentation. To our knowledge, this is one of the first studies to systematically integrate a ResNet-101 encoder with DeepLabv3 + and conduct an extensive backbone comparison (including ResNet-50, MobileNetV2, and EfficientNet-B0) on the ISIC-2016 dataset. Additionally, performance has been validated through learning curve analysis and confusion matrices, confirming both accuracy and generalization across lesion types. The results have highlighted the model’s novelty, its ability to address critical research gaps in multi-scale and boundary-aware segmentation, and its strong potential for clinical integration, offering a reliable and scalable solution for dermatological image analysis.