Automatic lesion border selection in dermoscopy images using morphology and color features.

Area of Science:

• Dermatology
• Computer Vision
• Medical Imaging

Background:

• Accurate skin lesion segmentation is crucial for computer-aided melanoma diagnosis.
• Variations in dermoscopy imaging techniques complicate automated skin lesion segmentation.
• Existing single segmentation algorithms often fail to provide acceptable lesion borders.

Purpose of the Study:

• To develop and evaluate a classifier for automatically selecting optimal skin lesion borders from multiple segmentation outputs.
• To enhance the accuracy of skin lesion border detection for improved computer-aided diagnosis (CADx) of melanoma.

Main Methods:

• A random forests classifier was trained to select the best lesion border from 12 different segmentation algorithms.
• The classifier utilized morphological and color features from regions inside and outside the candidate borders.
• A "good-enough" border criterion was used for evaluating segmentation quality.

Main Results:

• The random forests classifier achieved a satisfactory border prediction rate of 96.38% on an 802-lesion test set.
• This performance significantly outperforms the best single segmentation algorithm, which achieved 85.91% accuracy.
• The classifier demonstrated robust performance across diverse dermoscopy images.

Conclusions:

• The proposed classifier-based approach significantly improves automatic skin lesion border detection compared to individual segmentation methods.
• This automated border selection method offers a more reliable solution for processing skin lesions in CADx systems.
• The findings suggest a promising advancement in automated dermatological image analysis for melanoma detection.