Malignant melanoma detection by Bag-of-Features classification

Area of Science:

• Dermatology
• Computer Science
• Medical Imaging

Background:

• Malignant melanoma detection from epiluminescence microscopy images is crucial for early diagnosis.
• Automated methods can aid dermatologists in screening skin lesions.
• Accurate classification of skin lesions is challenging due to visual similarities.

Purpose of the Study:

• To develop and evaluate a Bag-of-Features approach for automated malignant melanoma detection.
• To compare the performance of Naive Bayes and Support Vector Machines for this classification task.
• To incorporate a model selection scheme that minimizes costly false negative errors.

Main Methods:

• A Bag-of-Features model was employed, representing each skin lesion by a histogram of visual codewords.
• Codewords were generated from a training dataset of epiluminescence microscopy images.
• Classification was performed using Naive Bayes and Support Vector Machines algorithms.
• The Neyman-Pearson score was utilized for model selection to penalize false negatives.

Main Results:

• The Bag-of-Features approach achieved a classification performance of 82.21% on a dataset of 100 skin lesion images.
• Support Vector Machines yielded better results compared to Naive Bayes in this context.
• The model selection scheme effectively addressed the asymmetry in error costs.

Conclusions:

• The Bag-of-Features method shows promise for malignant melanoma detection using epiluminescence microscopy.
• The chosen classification algorithms and model selection strategy are effective for this medical imaging task.
• Further research could explore larger datasets and advanced feature extraction techniques.