Segmentation of 3D MRI Using 2D Convolutional Neural Networks in Infants' Brain

Magnetic resonance (MR) imaging of the human brain poses significant challenges when it comes to segmenting the white matter, gray matter, and cerebrospinal fluid. This study presents a novel 2D model for segmenting 3D MR scans, utilizing 3D features, while maintaining a reduced number of model parameters compared to traditional 3D deep neural network models. The proposed method addresses the intensity contrast issue between white and gray matter in six- to nine-month-old infants by leveraging consecutive concatenation slices as a three-channel input image. Additionally, the combination of T1 and T2 weighted MR images for each patient reduces model complexity. Specifically, our study presents a 2D model capable of effectively segmenting MR images of the human brain, especially when there is a close contrast between white matter and gray matter. The combination of 3D features and a reduced parameter count improves segmentation accuracy. Our findings suggest the potential of our proposed method for the diagnosis of possible brain abnormalities. This will pave the way for more accurate and efficient medical image analysis in neuroimaging. To evaluate the effectiveness of our approach, an extensive evaluation was conducted on the iSeg-2017 datasets. The results demonstrated substantial improvements in the segmentation accuracy compared to other 2D techniques, especially in limited annotation settings. The proposed method achieved impressive Dice scores of 0.803, 0.817, and 0.907 for white matter, gray matter, and cerebrospinal fluid, respectively. Accordingly, these results demonstrate the efficiency of our 2D model in accurately segmenting brain tissue in MR images.