Training With Uncertain Annotations for Semantic Segmentation of Basal Cell Carcinoma From Full-Field OCT Images

Area of Science:

• Medical Imaging
• Dermatology
• Computer Vision

Background:

• Semantic segmentation of basal cell carcinoma (BCC) from full-field optical coherence tomography (FF-OCT) images is crucial for skin cancer diagnosis.
• Annotation uncertainty in FF-OCT data due to lack of color specificity challenges model training and impacts BCC segmentation performance.

Purpose of the Study:

• To develop an approach for training segmentation models with uncertain annotations for BCC detection in FF-OCT images.
• To enhance the accuracy and robustness of BCC segmentation by addressing annotation challenges and improving model performance.

Main Methods:

• Implemented a data selection strategy to mitigate annotation uncertainty.
• Expanded the segmentation classes to include sebaceous glands and hair follicles.
• Utilized a self-supervised pre-training procedure to optimize initial model weights.
• Developed three post-processing techniques to reduce speckle noise and image discontinuities.

Main Results:

• Achieved a mean Dice score of 0.503±0.003 for BCC segmentation.
• Demonstrated the best performance to date for semantic segmentation of BCC from FF-OCT images.

Conclusions:

• The proposed approach effectively addresses the challenge of uncertain annotations in BCC segmentation from FF-OCT images.
• The integration of data selection, class expansion, self-supervised pre-training, and post-processing techniques significantly improves segmentation performance.