Convolutional Neural Network (CNN) models often exhibit sub-optimal performance without labeled datasets. For diseases like brain tumors, the availability of labels is challenging, and obtaining scans for specific grades of tumors is even more difficult. To tackle these issues, this paper introduces an innovative approach that utilizes self-supervised learning (SSL) to learn insights from unlabeled data and few-shot learning to handle the small sample size problem. Current methods targeting small datasets yield poor, often over-segmented results, rendering them ineffective for clinical applications. Particularly for brain tumor MR images, we propose to learn more information via registration as a pretext task. We register deformed images using an encoder-decoder model, allowing us to learn a more adequate representation of the segmentation target. The trained model extracts prototypes from the support images, employing them as guidance to generate segmentation masks for the query images. The use of registration as a pretext task in the few-shot model sees a dice score of 75.18%, a 1.28% increase over the baselines for the BRATS2021 in the 1-way-1-shot setting. Furthermore, our method demonstrates a remarkable ability to delineate tumor boundaries. The Hausdorff Distance (HD) of the model drops to 10.03 from 14.29, showing the model's ability to identify tumor boundaries. This novel approach holds significant potential for addressing the scarcity of labeled data in MR imaging, particularly in scenarios where obtaining precise annotations is challenging. With this approach, we aim to equip clinicians with valuable tools for more precise and effective diagnosis and treatment of brain tumors.

Convolutional Neural Network (CNN) models often exhibit sub-optimal performance without labeled datasets. For diseases like brain tumors, the availability of labels is challenging, and obtaining scans for specific grades of tumors is even more difficult. To tackle these issues, this paper introduces an innovative approach that utilizes self-supervised learning (SSL) to learn insights from unlabeled data and few-shot learning to handle the small sample size problem. Current methods targeting small datasets yield poor, often over-segmented results, rendering them ineffective for clinical applications. Particularly for brain tumor MR images, we propose to learn more information via registration as a pretext task. We register deformed images using an encoder-decoder model, allowing us to learn a more adequate representation of the segmentation target. The trained model extracts prototypes from the support images, employing them as guidance to generate segmentation masks for the query images. The use of registration as a pretext task in the few-shot model sees a dice score of 75.18%, a 1.28% increase over the baselines for the BRATS2021 in the 1-way-1-shot setting. Furthermore, our method demonstrates a remarkable ability to delineate tumor boundaries. The Hausdorff Distance (HD) of the model drops to 10.03 from 14.29, showing the model's ability to identify tumor boundaries. This novel approach holds significant potential for addressing the scarcity of labeled data in MR imaging, particularly in scenarios where obtaining precise annotations is challenging. With this approach, we aim to equip clinicians with valuable tools for more precise and effective diagnosis and treatment of brain tumors.