Uncertainty Aware Temporal-Ensembling Model for Semi-Supervised ABUS Mass Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate breast mass segmentation in automated breast ultrasound (ABUS) is vital for 3D reconstruction and surgical planning.
• Deep learning, particularly convolutional neural networks (CNNs), shows promise but is hindered by limited annotated data.
• Existing methods struggle with unreliable predictions in segmentation models.

Purpose of the Study:

• To develop an uncertainty-aware temporal ensembling (UATE) model for semi-supervised ABUS mass segmentation.
• To enhance segmentation performance by addressing the issue of unreliable predictions in deep learning models.
• To leverage a large unlabeled dataset to improve segmentation accuracy with limited labeled data.

Main Methods:

• A temporal ensembling segmentation (TEs) model was designed for semi-supervised ABUS mass segmentation.
• An uncertainty map was estimated for each image to identify unreliable predictions.
• An adaptive ensembling momentum map and an uncertainty-aware unsupervised loss were integrated into the TEs model.

Main Results:

• The UATE model achieved a Jaccard index (JI) of 63.65%, Dice similarity coefficient (DSC) of 74.25%, accuracy (AC) of 99.21%, and Hausdorff distance (HD) of 3.81mm on the ABUS dataset.
• The semi-supervised UATE method demonstrated superior performance compared to the fully supervised approach.
• The proposed method showed promising results when compared to existing semi-supervised segmentation techniques.

Conclusions:

• The UATE model effectively improves semi-supervised ABUS mass segmentation by incorporating uncertainty estimation.
• The developed method offers a viable solution for leveraging unlabeled data to enhance segmentation accuracy in medical imaging.
• This approach holds potential for improving computer-aided diagnosis and surgical planning in breast imaging.