Comparison of Different Deep Learning Frameworks for Hippocampus Body and Head Segmentation

Abstract

Purpose: Precise hippocampus segmentation from Magnetic Resonance Imaging (MRI) scans is vital in diagnosing various neurological disorders. Traditional segmentation methods face challenges due to the hippocampus's complex structure. This study compares four deep learning frameworks to segment hippocampal parts, including concurrent, separated, ordinal, and attention-based strategies. Materials and Methods: This research utilized 3D T1-weighted MR images with manually delineated hippocampus head and body labels from 260 participants. The images were randomly split into five folds for experimentation, each time one of those was designated as the test set and the rest as the training set. Results: The findings indicate that both concurrent and separated frameworks perform better than the ordinal and attention-based frameworks regarding the Dice and Jaccard coefficients. In head segmentation, the separated framework had a Dice similarity of 0.8748, a Jaccard similarity of 0.7794, and a Hausdorff distance of 5.4160. In body segmentation, the concurrent framework had a Dice similarity of 0.8616, a Jaccard similarity of 0.7591, and a sensitivity of 0.8437. Statistical results from the one-way ANOVA test showed a significant difference in performance for the body part (P-value=0.008), but not for the head region (P-value=0.652) between concurrent and separated frameworks. Comparing the concurrent with ordinal and attention-based frameworks showed a significant difference in both body and head regions (P-value<0.001). Conclusion: Researchers must consider the differences between various frameworks while selecting a segmentation method for their specific task. Understanding the strengths and weaknesses of every framework is essential for deciding on the top-rated segmentation approach for precise applications.