Rethinking Skin Lesion Segmentation in a Convolutional Classifier.

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Melanoma is a dangerous skin cancer if not diagnosed early.
• Computer-aided diagnosis (CAD) systems using convolutional neural networks (CNNs) show promise for improving diagnostic accuracy.
• The role of image segmentation in CNN-based skin lesion classification is unclear, with prior studies showing conflicting results.

Purpose of the Study:

• To investigate the impact of enlarging the segmentation border of skin lesions on classification performance.
• To determine if expanding segmentation beyond the lesion boundary improves diagnostic accuracy compared to perfect segmentation or no segmentation.

Main Methods:

• Utilized a convolutional neural network (CNN) classifier within a transfer learning framework.
• Experimented with preprocessing techniques that enlarge the segmentation border to include surrounding skin pixels.
• Compared classification performance using enlarged segmentation borders against perfect segmentation (dermatologist-created masks) and no segmentation.

Main Results:

• Segmentation border enlargement, to a certain degree, improved classification performance across various metrics.
• The enlarged segmentation approach outperformed both perfect segmentation and no segmentation methods.
• This suggests that including surrounding non-lesion skin can benefit CNN-based classifiers.

Conclusions:

• Enlarging the segmentation border is a beneficial preprocessing step for CNN-based melanoma classification.
• This method shows potential to enhance diagnostic accuracy in computer-aided diagnosis systems.
• Preprocessing strategies that expand segmentation boundaries may offer superior performance over traditional segmentation approaches.