Feasibility Study of Detecting and Segmenting Small Brain Tumors in a Small MRI Dataset with Self-Supervised Learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate segmentation and detection of small metastatic brain tumors are crucial for effective treatment planning.
• Limited availability of annotated medical datasets poses a significant challenge for training deep learning models.

Purpose of the Study:

• To evaluate the feasibility of training a deep neural network for segmenting and detecting small metastatic brain tumors in MRI scans.
• To assess the effectiveness of using a small dataset (33 cases) by leveraging large public datasets of primary tumors.

Main Methods:

• Exploration of supervised learning, transfer learning (two approaches), and self-supervised learning (SSL).
• Utilized U-net and Swin UNETR deep learning models for tumor segmentation and detection.
• Leveraged large public datasets of primary tumors to augment the small dataset of metastatic tumors.

Main Results:

• Self-supervised learning with the Swin UNETR model demonstrated the best performance.
• Achieved a Dice score of approximately 0.19 for small brain tumors.
• Attained 100% sensitivity and 54.5% specificity, which improved to 80.0% when excluding subjects with hyperintensities.

Conclusions:

• It is feasible to train effective models for small brain tumor segmentation and detection using SSL with limited data.
• SSL combined with transfer learning shows promise for overcoming data scarcity in medical imaging AI.