An improved strategy for skin lesion detection and classification using uniform segmentation and feature selection...

Area of Science:

• Dermatology
• Medical Imaging
• Computer Science

Background:

• Melanoma, a severe skin cancer, necessitates early diagnosis for improved survival rates.
• Traditional diagnostic methods are resource-intensive, requiring expert interpretation and specialized equipment.
• Advancements in computational solutions offer promising alternatives for accurate and efficient skin lesion analysis.

Purpose of the Study:

• To propose and evaluate a novel computational method for classifying melanoma versus benign skin lesions.
• To integrate image preprocessing, segmentation, feature extraction, selection, and classification into a cohesive diagnostic system.

Main Methods:

• Image preprocessing included hair removal (DullRazor) and contrast enhancement using texture and color information.
• A hybrid lesion segmentation technique fused results via probability, followed by serial feature extraction (color, texture, HOG).
• Feature selection employed a novel Boltzman Entropy method, with final classification by Support Vector Machine.

Main Results:

• The proposed method achieved high performance on the PH2 dataset: 97.7% sensitivity, 96.7% specificity, 97.5% accuracy, and 97.5% F-score.
• These results significantly surpass existing methods evaluated on the same dataset.
• The system demonstrated superior detection and classification of melanoma compared to current approaches.

Conclusions:

• The developed computational method offers a highly accurate and efficient approach for melanoma diagnosis.
• This technique holds potential to augment traditional diagnostic procedures, improving early detection rates.
• The integration of advanced image processing and machine learning techniques shows significant promise in dermatological applications.