Enhancing Slice-Wise Brain MRI Tasks using Self-Supervised and Auxiliary Learning.

Area of Science:

• Medical Image Processing
• Deep Learning
• Artificial Intelligence

Background:

• Deep learning for medical imaging requires large labeled datasets, which are time-consuming and costly to acquire.
• Self-supervised learning (SSL) offers a solution by learning from unlabeled data, reducing the need for manual annotations.

Purpose of the Study:

• To evaluate SSL methods for craniopharyngioma recognition (CPGR) and detection of hypothalamic involvement (DHI) using brain MRI.
• To introduce and assess Slice-Wise Regularization (SWR) as a novel auxiliary task to enhance SSL performance on MRI data.

Main Methods:

• Compared supervised learning with SSL methods (SimCLR, DCL, VICReg) followed by fine-tuning on brain MRI slices.
• Introduced Slice-Wise Regularization (SWR), an annotation-free auxiliary task leveraging intrinsic MRI properties.
• Evaluated SSL + SWR against supervised learning for CPGR and DHI tasks.

Main Results:

• SSL + SWR achieved statistically significant improvements over supervised learning.
• Achieved F1-scores of 80.3 ± 2.4 for CPGR and 82.8 ± 5.0 for DHI with SSL + SWR.
• Supervised learning achieved F1-scores of 74.4 ± 4.9 for CPGR and 65.4 ± 6.5 for DHI.

Conclusions:

• SSL combined with SWR is a highly effective approach for medical image analysis tasks like CPGR and DHI.
• The proposed SWR method enhances SSL by preserving representations of contiguous MRI slices without additional annotations.
• This approach offers a promising direction for developing robust deep learning models in medical imaging with limited labeled data.