AmygdalaGo-BOLT：Accurate Segmentation of Amygdala Using a Boundary Contrastive Learning Transformer

1897 

Abstract Submission 

Bo Dong¹, Quan Zhou², Xi-Nian Zuo^3,4,5, Hongjian He^6,7

¹College of Biomedical Engineering and Instrument Science, Zhejiang University, Hangzhou, China, ²Faculty of Psychology, Beijing Normal University, Beijing, China, ³State Key Laboratory of Cognitive Neuroscience and Learning, Beijing Normal University, Beijing, China, ⁴National Basic Science Data Center, Beijing, Beijing, China, ⁵Institute of Psychology, Chinese Academy of Sciences, Beijing, China, ⁶School of Physics, Zhejiang University, Hangzhou, China, ⁷State Key Laboratory of Brain-Machine Intelligence, Zhejiang University, Hangzhou, China

Bo Dong  
College of Biomedical Engineering and Instrument Science, Zhejiang University
Hangzhou, China

Quan Zhou  
Faculty of Psychology, Beijing Normal University
Beijing, China

Xi-Nian Zuo  
State Key Laboratory of Cognitive Neuroscience and Learning, Beijing Normal University|National Basic Science Data Center, Beijing|Institute of Psychology, Chinese Academy of Sciences
Beijing, China|Beijing, China|Beijing, China

Hongjian He  
School of Physics, Zhejiang University|State Key Laboratory of Brain-Machine Intelligence, Zhejiang University
Hangzhou, China|Hangzhou, China

The amygdala is involved in various brain functions, and the accurate segmentation of this structure is is a crucial issue in cognitive neuroscience (Avecillas et al. 2023, Chai et al. 2023, Nguyen et al.2023). A major challenge lies in the small volume of the amygdala and the low contrast between the amygdala and the surrounding complex tissues, leading to uncertainty in boundary delineation. To address these issues, we introduced AmygdalaGo-BOLT, a 3D transformer architecture tailored for small structures, incorporating a boundary contrastive learning optimization algorithm. Compared to Freesurfer (Fischl et al. 2002) and Unet (Falk et al. 2019), the proposed method showed superior performance. Additionally, we validated our segmentation approach by applying it to assess volume changes during child development (Zhou et al. 2021).

The AmygdalaGo-BOLT, innovatively employs 3D transformer to construct the long-range relationships between voxels, thereby enhancing feature and enabling a more holistic comprehension. It is specifically executed by integrating the attention, which considers for the differences between the amygdala and other structures as well as the similarities within the amygdala, as opposed to merely focusing on local regions. Furthermore, to augment the alignment between predicted and actual boundaries, we implement an boundary contrastive learning optimization method. This optimization brings the edges closer to feature points that resemble the amygdala and distances those non-amygdala-like feature points in a high-dimensional feature space to amplify feature differences between the amygdala and surrounding tissue and reducing intra-amygdala variation.

We conducted a comprehensive comparison between our segmentation method and other existing approaches dedicated to the segmentation of the amygdala structure. Notably, our method outperforms both the classical Freesurfer and the widely-used Unet, achieving segmentation results that align closely with human-level performance as illustrated in Figure 2(a). Specifically, our dice score is close to 90%, a substantial improvement over Freesurfer and Unet. The superiority of our amygdala segmentation model is further evident in the precision metric, where we exhibit a significant lead, as illustrated in Figure 2(b). Figure 2(c) displays the growth curve generated by our algorithm, revealing a strong correlation with gender and age and surpassing the performance of alternative algorithms. This emphasizes the robustness and accuracy of our algorithm in estimating amygdala volume while accounting for variations in sex and age. By closely approximating distribution patterns of amygdala volume across different demographics, our algorithm offers valuable insights into developmental changes in the amygdala. This information enhances our understanding of how the amygdala evolves over time in relation to sex differences and age-related factors. In summary, the ability of our algorithm to precisely capture and represent the distribution of amygdala volume with respect to sex and age underscores its robustness and effectiveness in characterizing changes in amygdala development. This contribution significantly advances the field of neurodevelopmental research.

In conclusion, our proposed AmygdalaGo-BOLT achieves accurate amygdala segmentation, as validated across multiple independent datasets. This deep-learning approach with a transformer architecture significantly enhances efficiency, completing the segmentation in just 5 seconds, thus outperforming popular software like Freesurfer. Our targeted network design, tailored for small volume and complex boundary structures, could potentially be extended to other complex brain structure segmentation problems, given sufficient labeled training data. Overall, AmygdalaGo-BOLT offers a valuable tool for morphometric analysis in neuroscience research.

Methods Development ¹

Segmentation and Parcellation ²

Segmentation