Improving Automatic Melanoma Diagnosis Using Deep Learning-Based Segmentation of Irregular Networks.

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Deep learning models show high accuracy in melanoma diagnosis but still miss a significant minority of cases.
• Irregular pigment networks, visible via dermoscopy, are often present in missed melanomas.
• Current deep learning approaches may not fully leverage subtle dermoscopic features.

Purpose of the Study:

• To develop an improved melanoma diagnostic pipeline by integrating deep learning with conventional image analysis.
• To create and utilize an annotated irregular pigment network database for enhanced feature extraction.
• To increase the diagnostic accuracy and recall for malignant melanoma detection.

Main Methods:

• An annotated irregular pigment network dataset was created from 487 dermoscopic melanoma images (ISIC 2019).
• Transfer learning segmentation models (e.g., U-Net++) were trained to identify irregular networks.
• A pipeline fused deep learning outputs with hand-crafted features (color, texture, shape) from detected networks for classification.

Main Results:

• The combined approach improved melanoma vs. benign recall by 11% and accuracy by 2% compared to deep learning alone.
• The random forest algorithm showed the highest recall improvement within the sequential pipeline.
• U-Net++ successfully generated irregular network masks for feature extraction.

Conclusions:

• Fusing deep learning with conventional image processing of dermoscopic features significantly enhances melanoma diagnostic accuracy.
• This hybrid approach leverages the strengths of both AI and traditional methods for improved clinical outcomes.
• Further research into combining automated dermoscopic feature detection with deep learning is recommended.