Supervised Saliency Map Driven Segmentation of Lesions in Dermoscopic Images

Area of Science:

• Medical Image Analysis
• Computational Dermatology
• Machine Learning for Healthcare

Background:

• Accurate lesion segmentation is crucial for automated melanoma recognition.
• Dermoscopic image challenges include color variations, occlusions, and artifacts, complicating segmentation.
• Existing methods struggle with these inherent image deficiencies.

Purpose of the Study:

• To develop an improved supervised saliency detection method (mDRFI) for robust lesion segmentation in dermoscopic images.
• To enhance lesion detection despite image quality issues.
• To create a comprehensive segmentation framework for melanoma recognition.

Main Methods:

• Proposed a discriminative regional feature integration (DRFI) method incorporating multilevel segmentation, regional contrast, and background descriptors.
• Introduced an enhanced method, mDRFI, with improved regional property and background descriptors using a novel thresholding algorithm.
• Integrated saliency maps into a level set framework for precise lesion boundary fitting.

Main Results:

• mDRFI demonstrated superior performance over DRFI in identifying lesions as salient objects.
• The proposed framework successfully generated initial lesion masks via thresholding and postprocessing.
• Evaluation on three public datasets showed the method outperforms conventional algorithms and rivals deep learning approaches.

Conclusions:

• The mDRFI method significantly enhances lesion segmentation accuracy in challenging dermoscopic images.
• The developed framework provides a robust and competitive solution for melanoma recognition.
• This approach holds promise for improving diagnostic accuracy in dermatological imaging.