Sammendrag
This work initiates a novel exploration of domain transfer learning in pre-clinical image analysis, concentrating on the segmentation of brain lesion areas of mice MR images. Two transfer learning techniques, partially fine-tuning and fully fine-tuning the model were used with the pre-trained RatLesNetV2 model. Various training samples ranging from 8 to 40 images, both with and without data augmentation, were used to further evaluate these techniques. Utilizing MONAI framework, the research also assessed the 3D UNet, UNetR, and Dynamic Unet models. When compared to the others, the 3D UNet with augmentation performed most effectively, with an average Dice Coefficient (DC) of 0.88 and 95% Hausdorff distance (HD95) of 0.39 mm. Partial fine-tuning produced segmentations with an average DC of 0.85 and HD95 of 0.71 and these segmentations were approximately similar to the ones generated by 3D UNet and better than the segmentations produced by other MONAI models, indicating the potential adaptability of partial fine-tuning. Full fine-tuning provided insightful results but did not outperform the other approaches used in this study, demonstrating the intricate interaction between overfitting and adaptability in transfer learning. This study is one of the first to examine the possibility of transfer learning in pre-clinical settings, bringing an important approach to a different field. The results provide a solid platform for understanding the underlying mechanisms that led to these findings and expanding the scope of transfer learning's applications in pre-clinical image analysis.