Fully Convolutional Neural Networks to Detect Clinical Dermoscopic Features.

Area of Science:

• Dermatology
• Computer Vision
• Medical Imaging

Background:

• Dermoscopic features in skin lesions can indicate melanoma, but manual detection is subjective.
• Automated detection offers potential for quantitative and reproducible melanoma diagnosis.

Purpose of the Study:

• To develop and evaluate a fully convolutional neural network (CNN) for segmenting clinical dermoscopic features in skin lesions.
• To address challenges in automated feature detection, including imbalanced datasets and evaluation metrics.

Main Methods:

• A fully convolutional neural network (CNN) architecture was designed using interpolated feature maps from intermediate layers.
• The model was trained to minimize a negative multilabel Dice-F1 score to handle imbalanced labels.
• The approach was evaluated on the 2017 ISIC-ISBI Dermoscopic Feature Classification Task dataset.

Main Results:

• The proposed CNN achieved first place in the 2017 ISIC-ISBI challenge, with an area under the receiver operator characteristic curve of 0.895.
• Simple baseline models were shown to outperform state-of-the-art methods under specific challenge metrics.
• A modified fuzzy Jaccard Index was proposed to better evaluate model performance, excluding empty sets.

Conclusions:

• Classifying clinical dermoscopic features can be effectively treated as a segmentation problem.
• Current challenge metrics may not fully capture the efficacy of segmentation models for dermoscopic feature detection.
• The study highlights the potential of deep learning for objective melanoma diagnosis and proposes improvements in evaluation methodologies.