Melanoma segmentation using deep learning with test-time augmentations and conditional random fields.

Area of Science:

• Medical Image Analysis
• Computer-Aided Diagnosis
• Dermatology

Background:

• Skin lesion segmentation is crucial for computer-aided diagnosis (CAD) but challenged by variations in lesion shape and size.
• Current deep learning segmentation methods lack the inter-observer agreement required for clinical assessment.
• Automated segmentation is needed to improve the accuracy and reliability of skin lesion quantification.

Purpose of the Study:

• To propose a novel, fully automated deep learning approach for accurate skin lesion segmentation.
• To enhance segmentation accuracy through sophisticated preprocessing and postprocessing techniques.
• To develop a robust framework for computer-aided skin lesion diagnostic systems.

Main Methods:

• Utilized UNet, ResUNet, and ResUNet++ deep learning models for segmentation.
• Implemented a preprocessing phase combining morphological filters and inpainting to remove hair artifacts.
• Applied test time augmentation (TTA) and conditional random field (CRF) for postprocessing refinement.

Main Results:

• Achieved high Jaccard Index scores on ISIC-2016 (up to 90.02%) and ISIC-2017 (up to 85.96%) datasets.
• Demonstrated improved segmentation accuracy with combined training on both datasets.
• The proposed method shows robustness and scalability for automated diagnostic systems.

Conclusions:

• The novel deep learning framework provides accurate and automated skin lesion segmentation.
• The sophisticated pre- and postprocessing steps enhance overall segmentation performance.
• This approach offers a scalable and robust solution for developing advanced computer-aided skin lesion diagnostic systems.