TMA-UNet: A Parameter-Efficient Deep Learning Model for Fetal Brain MRI Segmentation

The segmentation of fetal brain structures in MRI images is crucial for the early diagnosis of neurological abnormalities. However, existing deep learning models suffer from high computational complexity due to the excessive number of learnable parameters, which limits their practical use in clinical settings. This paper introduces TMA-UNet, a novel deep learning architecture designed to reduce the parameter count while maintaining high segmentation accuracy. Experimental results on the Fetal Brain Atlas from Harvard University demonstrate that our model outperforms other models, achieving an average Dice coefficient of 99.02%, IoU of 98.85%, and F1-score of 99.04%. Notably, TMA-UNet achieves this performance with only 1.39 million parameters, a substantial reduction compared to other state-of-the-art segmentation models. This efficiency enhances the model’s feasibility for real-world deployment and establishes a new benchmark for parameter-efficient medical image segmentation.