Skin Lesion Segmentation in Dermoscopic Images with Noisy Data

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate skin lesion segmentation is vital for diagnosing skin cancer.
• Existing deep learning models face challenges with noisy ground truth data in dermoscopic datasets.

Purpose of the Study:

• To develop and evaluate a deep learning model for skin lesion segmentation.
• To investigate the impact of noisy labels on model training and performance evaluation.

Main Methods:

• A novel network architecture utilizing EfficientNet and squeeze-and-excitation residual structures.
• Application on the International Skin Imaging Collaboration (ISIC) 2017 dataset.
• Manual curation of ground truth labels into good, mildly noisy, and noisy categories.

Main Results:

• The proposed model achieved Jaccard scores of 0.807 (official test set) and 0.832 (curated test set).
• Noisy training labels did not significantly degrade segmentation performance.
• Noisy test labels negatively impacted evaluation scores, highlighting the need for clean test data.

Conclusions:

• The developed deep learning approach demonstrates superior performance in skin lesion segmentation.
• The presence of noisy labels in the test set can lead to inaccurate performance assessments.
• Future studies should prioritize using curated, noise-free test datasets for robust algorithm evaluation.