Melanoma Recognition in Dermoscopy Images via Aggregated Deep Convolutional Features

Area of Science:

• Dermatology
• Computer Vision
• Machine Learning

Background:

• Accurate melanoma recognition is crucial for early diagnosis and treatment.
• Existing methods face challenges with intra-class variations and inter-class similarities in dermoscopy images.
• Limited training data often hinders the performance of deep learning models.

Purpose of the Study:

• To develop a novel framework for dermoscopy image recognition using a combination of deep learning and local descriptor encoding.
• To improve the discriminative power of features for classifying melanoma.
• To address challenges posed by limited training data in skin lesion classification.

Main Methods:

• Extracted deep representations using a very deep residual neural network pretrained on natural images.
• Aggregated local deep descriptors using Fisher vector (FV) encoding for global image representation.
• Classified melanoma images using a support vector machine with a Chi-squared kernel on FV encoded features.

Main Results:

• The proposed method generated more discriminative features, effectively handling variations within melanoma classes and between melanoma and non-melanoma classes.
• Achieved superior performance compared to state-of-the-art methods on the ISBI 2016 Skin Lesion Challenge dataset.
• Demonstrated effectiveness even with limited training data.

Conclusions:

• The novel framework combining deep learning and FV encoding offers a robust approach for dermoscopy image recognition.
• The method shows significant potential for improving melanoma diagnosis accuracy.
• This approach provides a valuable contribution to the field of automated skin lesion analysis.