Segmentation of dermoscopy images based on deformable 3D convolution and ResU-NeXt +

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Dermatology

Background:

• Current melanoma segmentation models like FCNs and U-Net suffer from parameter redundancy and vanishing gradients.
• These limitations reduce the Jaccard index, impacting the accuracy of skin lesion image segmentation.

Purpose of the Study:

• To improve melanoma patient survival rates by enhancing skin lesion segmentation accuracy.
• To address the limitations of existing neural network models in segmenting skin lesions.

Main Methods:

• Proposed an improved skin lesion segmentation model: D3DC-ResU-NeXt++, incorporating deformable 3D convolution and ResU-NeXt++.
• Introduced a new data preprocessing method (DCRH) involving dilation, crop, resizing, and hair removal.
• Utilized rectified Adam (RAdam) as the training optimizer for faster convergence and to avoid local optima.

Main Results:

• The D3DC-ResU-NeXt++ model demonstrated excellent performance on the ISIC2018 Task I dataset.
• Achieved a Jaccard index of 86.84%, significantly improving skin lesion image segmentation.
• The model effectively highlights lesion areas and improves segmentation robustness.

Conclusions:

• The proposed D3DC-ResU-NeXt++ model enhances skin lesion segmentation accuracy and efficiency.
• This advancement can assist dermatologists in diagnosing skin lesions and improving patient outcomes.
• The method contributes to improving the survival rates of skin cancer patients.