Improving Skin Lesion Segmentation with Self-Training.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Skin lesion segmentation is crucial for accurate skin cancer diagnosis.
• High-quality segmentation requires extensive labeled data, which is costly and time-consuming to acquire.
• Semi-supervised learning offers a solution by leveraging unlabeled data to improve model performance.

Purpose of the Study:

• To propose a novel semi-supervised approach for skin lesion segmentation using self-training with a Noisy Student.
• To reduce the reliance on manually annotated data in medical image segmentation.
• To enhance the accuracy of skin cancer diagnosis through improved segmentation.

Main Methods:

• Implemented a semi-supervised learning framework utilizing the Noisy Student strategy.
• Employed the DeepLabV3 architecture for both teacher and student models.
• Utilized a four-step process involving teacher training, pseudo-label generation, and iterative student training.

Main Results:

• Achieved a mean Intersection over Union (mIoU) of 88.0% on the ISIC 2018 dataset.
• Attained a mIoU of 87.54% on the PH2 dataset.
• Demonstrated improved segmentation performance with limited labeled data.

Conclusions:

• The proposed Noisy Student training approach effectively enhances neural network performance for skin lesion segmentation.
• This method significantly mitigates the need for extensive manual annotation, making it a valuable tool for medical image analysis.
• The approach shows promise for improving the accuracy and efficiency of skin cancer diagnosis.