Automated Deep Learning for Locus Coeruleus Segmentation in Aging and Parkinson's Disease

Overview

This study presents a novel deep learning pipeline for automated localization and segmentation of the locus coeruleus (LC) using neuromelanin-sensitive MRI. The method demonstrates improved precision and efficiency over previous approaches and generalizes well to Parkinson's disease (PD) subjects without fine-tuning. Automated extraction of contrast ratios (CRs) from segmented LC and pons regions facilitates objective biomarker analysis.

Background

The locus coeruleus (LC) is a small brainstem nucleus critically involved in neurodegenerative diseases such as Alzheimer's and Parkinson's disease. Neuromelanin-sensitive MRI enables in vivo visualization of the LC by exploiting its neuromelanin content. However, LC segmentation is challenging due to its small size and low MRI resolution, resulting in low inter-rater agreement among experts. Automated, reliable segmentation methods are needed to improve biomarker extraction and comparability across studies.

Data Highlights

Key Findings

• Proposed an iterative, multi-scale LC localization network (CoRe-Unet) that improves precision, reduces GPU memory use, and shortens training time compared to prior methods.
• Utilized multi-rater manual segmentations during training, which enhanced segmentation performance over single-rater training.
• Demonstrated that the pipeline trained on healthy aging data generalizes effectively to Parkinson's disease subjects without additional fine-tuning.
• Implemented fully automated LC biomarker extraction by segmenting the pons with a 3D-Unet and applying robust post-processing to calculate contrast ratios.
• Achieved higher segmentation accuracy than previous atlas-based and classical methods, which had Dice similarity coefficients around 0.4–0.45, with expert inter-rater agreement ranging from 0.499 to 0.68.

Clinical Implications

This automated deep learning pipeline enables objective and reproducible segmentation of the LC and related brainstem structures, facilitating consistent biomarker extraction in both healthy aging and Parkinson's disease populations. The approach may accelerate research into LC involvement in neurodegeneration and support early diagnosis and monitoring by providing reliable imaging-based markers. Reduced reliance on manual segmentation also improves scalability and comparability across studies.

Conclusion

The study introduces a robust, efficient deep learning framework for LC localization and segmentation using neuromelanin-sensitive MRI, demonstrating strong performance and generalizability. This advancement supports the development of imaging biomarkers critical for understanding and managing neurodegenerative diseases.

References

1. Hämmerer et al. 2018 -- Neuromelanin-sensitive MRI of the locus coeruleus
2. Dahl et al. 2020 -- Multi-rater segmentation improves LC segmentation
3. Sulzer et al. 2018 -- Neuromelanin imaging in Parkinson's disease
4. UK Parkinson’s Disease Brain Bank criteria 1992 -- Diagnostic criteria for PD
5. CoRe-Unet architecture and DSNT layer 2019 -- Coordinate regression in CNNs