Machine learning based skin lesion segmentation method with novel borders and hair removal techniques

Area of Science:

• Dermatology
• Medical Imaging
• Computer Science

Background:

• Accurate skin lesion segmentation is crucial for computer-aided diagnosis (CAD) of skin cancer.
• Existing segmentation methods struggle with artifacts in dermoscopic images, such as hair, varying borders, and low contrast.
• These artifacts reduce the diagnostic accuracy of CAD systems.

Purpose of the Study:

• To develop an effective skin lesion segmentation method for dermoscopic images, addressing limitations of current techniques.
• To improve the performance of CAD systems by enhancing the accuracy of skin lesion identification.
• To accurately segment skin lesions despite the presence of challenging artifacts.

Main Methods:

• A novel method was proposed to first detect and remove artifacts like corner borders and hairs from dermoscopic images.
• The artifact-removed images were then enhanced using a state-of-the-art image enhancement technique.
• Lesion segmentation was performed using the GrabCut machine learning algorithm.

Main Results:

• The proposed method achieved a Jaccard Index of 0.77 on the PH2 dataset and 0.80 on the ISIC 2018 dataset.
• The Dice Index scores were 0.87 for PH2 and 0.82 for ISIC 2018.
• These results indicate superior performance compared to existing state-of-the-art skin lesion segmentation techniques.

Conclusions:

• The proposed method effectively segments skin lesions from dermoscopic images, even in the presence of artifacts.
• The technique shows significant improvements over existing methods, paving the way for more reliable CAD systems.
• This advancement can lead to earlier and more accurate diagnosis of skin cancer.