2D U-Net Framework for Multiclass Automated Segmentation of Healthy Structures, Ischemic Stroke, and Glioblastoma in the Rat Brain.
Authors
Affiliations (3)
Affiliations (3)
- Neuroimaging and Biotechnology Laboratory (NOBEL), Clinical Neurosciences Research Laboratory (LINC), Health Research Institute of Santiago de Compostela (IDIS), Santiago de Compostela, Spain.
- Department of Functional Biology (Physiology), University of Oviedo, Oviedo, Spain.
- Neuroradiology Group, Health Research Institute of Santiago de Compostela (IDIS), Santiago de Compostela, Spain.
Abstract
Rodents are essential for preclinical brain disease research, including stroke, tumors, and neurodegenerative disorders. Magnetic resonance imaging (MRI) enables longitudinal brain tissue visualization, but accurate segmentation is crucial for meaningful analysis. This study introduces a multitask 2D U-Net, a convolutional neural network designed for automated segmentation and quantification of brain regions, ischemic stroke lesions, and tumors in rats using T<sub>2</sub>-weighted imaging (T<sub>2</sub>-wi) and T<sub>2</sub>-relaxation maps. The model was evaluated on the following three MRI datasets: (i) segmentation of multiple brain regions in healthy rats, (ii) ischemic stroke lesions using T<sub>2</sub>-relaxation maps, and (iii) glioblastoma tumors using T<sub>2</sub>-wi. Performance was assessed using the Dice coefficient (DC), Hausdorff distance (HD), sensitivity (SEN), and volume ratio (VR), comparing network outputs against semiautomatic ground truths and existing models. A total of 1949 images from 111 rats were analyzed, expanded to 5847 through data augmentation. After correcting an animal-level train/test partitioning that prevented data leakage between subjects, the model achieved robust accuracy across all tasks (DC = 0.80-0.96, HD = 0.83-2.16 mm, SEN = 0.75-0.96, VR = 0.89-1.10), showing a strong correlation ( <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>ρ</mi> <mo>=</mo> <mn>0.82</mn> <mo>-</mo> <mn>0.99</mn></mrow> </math> ) with manual annotations. This framework improves segmentation accuracy, robustness, and applicability across different MRI settings and conditions, addressing issues like interrater variability and time inefficiency when analyzing rodent brain MRI data. Datasets are publicly available at https://github.com/NOBEL-MRI/Rat-Datasets.