Enhancing Hippocampal Subfield Visualization Through Deep Learning Reconstructed MRI Scans

Abstract

Background/Objectives: Assessing hippocampal pathology in epilepsy is challenging, and improving diagnostic accuracy can benefit from deep learning image reconstruction, standardized imaging protocols, and advanced post-processing methods. This study compares T2 TSE DRB (Deep Resolve Boost) sequences with standard T2 TSE sequences for hippocampal segmentation and volumetry using FreeSurfer, focusing on how DRB affects image acquisition time without compromising diagnostic accuracy. Methods: FreeSurfer (version 7.4.1) was used to segment hippocampal subregions in 36 subjects (mean age of 39 ± 14 years; 21 males, 15 females) using both T2 TSE DRB and T2 TSE sequences. The segmented volumes were compared with a two-tailed t-test, and pathological volume differences were assessed using z-values based on a 95% confidence interval (-2 < z < 2). Results: Overall hippocampal segment volumes were identical between sequences. However, significant volume differences were noted in the CA1-Body (p = 0.003), CA4-Body (p = 0.002), and whole hippocampal body (p = 0.012) in the right hippocampus. Despite these differences, the low effect sizes suggest DRB sequences are comparable to conventional sequences. Additionally, DRB reduced image acquisition time by 61%. Z-scores identified pathological volume changes between the left and right hippocampus in individual subjects. Conclusions: T2 TSE DRB sequences are non-inferior to conventional T2 TSE sequences for hippocampal segmentation. The DRB method improves efficiency while providing clinically reliable results, and the proposed 95% confidence interval can aid in more objective assessments of hippocampal pathology.

Keywords: FreeSurfer; deep learning; epilepsy; hippocampus; magnetic resonance tomography.

Figure 1

K-space data and coil sensitivity maps are input into a convolutional neural network (CNN). The CNN reconstructs an image, which is then compared to a high-quality reference. The resulting error (image output difference) is iteratively fed back into the network to refine the reconstruction. After training converges, a final image is produced with fixed model parameters [5,6,7]. * Deep Learning.

Figure 2

Hippocampal segmentation using FreeSurfer (healthy subject). The figure displays segmented hippocampal regions within a healthy individual using FreeSurfer. The left side shows images before segmentation, while the right side shows images after segmentation. The top row presents images obtained with the conventional T2 TSE sequence, while the bottom row presents images obtained with the T2 TSE DRB sequence. Hippocampal regions by color: parasubiculum, HATA, fimbria, hippacampal_fissure, HP_tail, presubiculum-head, presubiculum-body, subiculum-head, subiculum-body, CA1-head, CA1-body, CA3-head, CA3-body, CA4-head, CA4-body, GC-ML-DG-head, GC-ML-DG-body, molecular_layer_HP-head, moleculcular_layer_HP-body.

Figure 3

Hippocampal segmentation using FreeSurfer (HS). The figure displays segmented hippocampal regions in a patient with right-sided hippocampal sclerosis (HS), processed using FreeSurfer. The left side shows images before segmentation, while the right side shows images after segmentation. The top row presents images obtained using the conventional T2 TSE sequence, and the bottom row presents images obtained using the T2 TSE DRB sequence. Hippocampal regions by color: parasubiculum, HATA, fimbria, hippacampal_fissure, HP_tail, presubiculum-head, presubiculum-body, subiculum-head, subiculum-body, CA1-head, CA1-body, CA3-head, CA3-body, CA4-head, CA4-body, GC-ML-DG-head, GC-ML-DG-body, molecular_layer_HP-head, moleculcular_layer_HP-body.

Figure 4

Overview of the workflow for hippocampal subregion segmentation and analysis using FreeSurfer. The process includes MRI acquisition, image co-registration, segmentation, manual quality control, and volume comparison between T2 TSE and T2 TSE DRB sequences using the volumes of hippocampal segments.