A Novel Framework for Melanoma Lesion Segmentation Using Multiparallel Depthwise Separable and Dilated Convolutions...

Area of Science:

• Dermatology and Medical Imaging
• Artificial Intelligence in Healthcare

Background:

• Skin cancer, particularly melanoma, has a low survival rate, making early diagnosis critical.
• Accurate segmentation of melanoma lesions is challenging due to visual similarities and intraclass variations.
• Existing traditional segmentation algorithms often lack automation and human input, limiting their clinical application.

Purpose of the Study:

• To develop an improved, automated segmentation model for melanoma lesions.
• To address the limitations of existing methods in accurately segmenting visually similar and varied lesions.

Main Methods:

• Proposed an improved segmentation model utilizing depthwise separable convolutions.
• Implemented parallel multidilated filters to enhance feature encoding and filter receptive fields.
• Evaluated the model on three diverse datasets: DermIS, DermQuest, and ISIC2016.

Main Results:

• Achieved a high Dice score of 97% for the DermIS and DermQuest datasets.
• Obtained a Dice score of 94.7% on the ISIC2016 dataset.
• Demonstrated superior performance in segmenting melanoma lesions compared to existing approaches.

Conclusions:

• The proposed depthwise separable convolution-based model significantly improves melanoma lesion segmentation accuracy.
• This automated approach offers a promising tool for early and accurate diagnosis of skin cancer.
• The model's effectiveness across multiple datasets suggests its potential for broad clinical utility.