Related Experiment Video
Updated: Jan 7, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.3K
LoRA-based methods on Unet for transfer learning in aneurysmal subarachnoid hematoma segmentation
Cristian Minoccheri1, Matthew Hodgman2, Haoyuan Ma3
1Gilbert S. Omenn Department of Computational Medicine and Bioinformatics, University of Michigan, 1109 Geddes Avenue, Arbor, MI, 48104, USA. minoc@umich.edu.
BMC Medical Imaging
|December 26, 2025
Summary
Transfer learning with Low-Rank Adaptation (LoRA) methods significantly improves automated segmentation of aneurysmal subarachnoid hemorrhage (SAH) compared to standard fine-tuning. Novel DoRA variants show superior accuracy, especially for small hemorrhages, enabling better diagnostic tools.
Area of Science:
- Medical image analysis and deep learning.
- Neurological emergency diagnostics.
- Computational neuroimaging.
Background:
- Aneurysmal subarachnoid hemorrhage (SAH) is a critical condition with high mortality.
- Automated SAH segmentation using deep learning faces data scarcity and generalization issues.
- Transfer learning and parameter-efficient methods like LoRA are underexplored for SAH.
Purpose of the Study:
- To evaluate transfer learning from traumatic brain injury (TBI) CT scans for aneurysmal SAH segmentation.
- To investigate the efficacy of Low-Rank Adaptation (LoRA) methods, including novel CP-LoRA and DoRA variants, for SAH segmentation.
- To compare these methods against standard fine-tuning on a Unet architecture.
Main Methods:
- Implemented a Unet model pre-trained on multi-institutional TBI CT data.
- Fine-tuned the model on aneurysmal SAH data using 3-fold cross-validation.
- Developed and compared novel CP-LoRA and DoRA variants against existing LoRA methods and standard fine-tuning.
Main Results:
- All transfer learning approaches outperformed no fine-tuning (mean Dice 0.410).
- LoRA-based methods, particularly DoRA-C at rank 64, achieved the highest performance (Dice 0.572).
- Segmentation accuracy improved with larger hemorrhage volumes; higher ranks (64-96) enhanced performance.
Conclusions:
- Transfer learning between hematoma types is feasible and effective for SAH segmentation.
- LoRA-based methods, especially DoRA variants, significantly outperform conventional Unet fine-tuning.
- Novel methods like CP-LoRA offer parameter efficiency, and higher ranks may benefit clinical applications.
