Medical image segmentation is a fundamental challenge in image processing and plays a crucial role in the eva‎luation and diagnosis of various diseases. In the domain of brain MRI images, precise segmentation of different brain structures, especially in fetal brains, is of particular significance. This process can significantly impact the diagnosis of brain disorders and the prediction of normal or abnormal brain development. Since medical images often exhibit diverse characteristics across patients, the development of effective methods for segmenting these images remains one of the major challenges. In this context, the primary objective of this research is to develop a deep learning-based model capable of accurately segmenting fetal brain structures, particularly the white matter and the cortical plate, from MRI images. To achieve this goal, a novel method for extracting 2D slices from 3D volumetric fetal brain images is proposed. This approach, referred to as "adaptive slicing," extracts 2D slices based on the size and specific features of the fetal brain. Compared to conventional fixed slicing methods, this technique ensures that important segmentation information is preserved in each slice. Furthermore, spatial dependencies in the images are reduced, which is especially significant in fetal MRI images that may be affected by fetal movement during imaging. This study introduces two deep learning methods for binary and multiclass segmentation with the aim of reducing learnable parameters while maintaining performance. Both methods share a similar backbone. A comparison of these methods with standard segmentation models demonstrates that both outperform the models being compared. By reducing the number of parameters and prediction time, our models have been shown to perform effectively in clinical tests and deliver higher accuracy. Compared to the baseline method, the first method achieved an average increase of two percentage points across all parameters, while the second proposed method achieved an average increase of one percentage point across all parameters. Additionally, both models achieved a 50% reduction in the number of learnable parameters compared to the baseline, resulting in a significant decrease in the test time for each sample.

نادر كريمي

شادرخ سماوي

محمدعلي خسروي فرد , محمدرضا احمدزاده