A Novel Approach to Skin Lesion Segmentation: Multipath Fusion Model with Fusion Loss.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Dermatology

Background:

• Skin lesion segmentation is crucial for early skin cancer detection.
• Artifacts like hair and poor contrast pose significant challenges in segmentation.
• Deep convolutional neural networks offer promising solutions for computer-aided diagnosis.

Purpose of the Study:

• To develop an advanced deep learning fusion model for automatic skin lesion segmentation.
• To introduce a novel fusion loss function to enhance model robustness.
• To improve the accuracy and reliability of skin cancer diagnostic systems.

Main Methods:

• A U-Net (U-Net + ResNet 2D) architecture was fused with 2D volumetric convolutional neural networks.
• A new fusion loss function combining Dice Loss (DL) and Focal Tversky Loss (FTL) was proposed.
• The model was trained and evaluated on a dataset of 2594 skin lesion images.

Main Results:

• The fused model achieved a Jaccard score of 0.837 and a Dice score of 0.918 on internal test data.
• On the ISIC 2018 Task 1 test set, the model obtained a Jaccard index of 0.800 and a Dice score of 0.880.
• The proposed fusion loss function demonstrated increased model robustness.

Conclusions:

• The developed deep learning fusion model significantly advances skin lesion segmentation accuracy.
• The novel fusion loss function enhances the robustness of the segmentation model.
• This approach surpasses existing methods, offering a more effective tool for dermatological diagnosis.