LoRA-PT: Low-rank adapting UNETR for hippocampus segmentation using principal tensor singular values and vectors

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Computational Neuroscience

Background:

• Accurate hippocampus segmentation is crucial for understanding psychiatric disorders.
• Deep learning models offer advanced segmentation but require extensive resources and data.
• Medical image segmentation often faces challenges with limited labeled training data.

Purpose of the Study:

• To develop a parameter-efficient fine-tuning (PEFT) method for hippocampus segmentation.
• To adapt a pre-trained UNETR model from the BraTS2021 dataset for hippocampus segmentation.
• To overcome the limitations of computational cost and data scarcity in deep learning for medical imaging.

Main Methods:

• Proposed LoRA-PT, a novel PEFT technique.
• Utilized tensor singular value decomposition to create low-rank tensors from transformer parameter matrices.
• Fine-tuned only the low-rank tensors, keeping residual tensors fixed.

Main Results:

• LoRA-PT demonstrated superior segmentation accuracy compared to existing PEFT methods.
• Significantly reduced the number of parameter updates required for fine-tuning.
• Validated on three public hippocampus datasets.

Conclusions:

• LoRA-PT offers an efficient and effective solution for hippocampus segmentation.
• The method successfully transfers knowledge from large datasets to tasks with limited data.
• LoRA-PT advances the application of deep learning in neuroimaging analysis.