Abstract
Tissue characterisation with Cardiovascular Magnetic Resonance (CMR) parametric mapping has the potential to detect and quantify both focal and diffuse alterations in myocardial structure not assessable by late gadolinium enhancement. Native T1 mapping, in particular, has shown promise as a useful biomarker to support diagnostic, therapeutic, and prognostic decision-making in ischaemic and non-ischaemic cardiomyopathies. This study presents an automated framework for tissue characterisation from native ShMOLLI T1 mapping at 1.5T using a Probabilistic Hierarchical Segmentation (PHiSeg) network. We leverage the uncertainty information provided by the Bayesian inference in a novel automated quality control (QC) step to identify uncertain T1 values. The pipeline was validated against manual analysis on a cohort of the UK Biobank containing healthy subjects and chronic cardiomyopathy patients. T1 values from automatic and manual segmentations were highly correlated (r=0.97), and the proposed method was used to derive reference T1 ranges from 14,683 CMR exams—the largest cohort to date using T1-mapping images.
Introduction
Native T1 mapping is a powerful, non-invasive imaging technique that allows for the quantitative assessment of the myocardium without the need for gadolinium-based contrast agents. By measuring the longitudinal relaxation time of myocardial tissue, clinicians can detect diffuse pathologies such as fibrosis, edema, and inflammation. However, the clinical and research utility of T1 mapping is often bottlenecked by the requirement for precise manual segmentation of the left ventricular (LV) myocardium.
Manual segmentation is not only time-consuming but also subject to inter-observer variability. While standard deep learning models (like U-Nets) can automate this process, they often suffer from "silent failures"—producing plausible-looking but anatomically incorrect segmentations when faced with image artifacts or poor contrast. To address this, we developed a fully automated pipeline that incorporates Bayesian uncertainty to perform robust quality control (QC), ensuring that only high-quality data informs clinical conclusions.
The Automated Pipeline
1. Bayesian Segmentation with PHiSeg
At the core of our framework is the Probabilistic Hierarchical Segmentation (PHiSeg) network. Unlike deterministic models, PHiSeg utilizes a conditional variational autoencoder (cVAE) approach to model the posterior distribution of segmentations. By sampling multiple plausible segmentations for a single input image, the network provides both a final consensus mask and a pixel-wise uncertainty map. This uncertainty map highlights regions where the model is less confident, often corresponding to ambiguous tissue boundaries or image artifacts.
2. Two-Stage Quality Control
To prevent erroneous data from contaminating large-scale analyses, we implemented a dual-stage QC process:
- ELBO Thresholding: The Evidence Lower Bound (ELBO) score from the PHiSeg network is used as an initial filter. A low ELBO score indicates that the input image is an outlier relative to the training distribution, suggesting a likely segmentation failure.
- VGG-16 Classifier: The original T1 map, the predicted segmentation mask, and the uncertainty map are concatenated and fed into a VGG-16 convolutional neural network. This classifier is specifically trained to distinguish between "accurate" and "inaccurate" segmentations, providing a final layer of automated oversight.
3. T1 Analysis and Blood Correction
Following segmentation and QC, the system extracts T1 values for the LV myocardium. To account for physiological variations such as hematocrit levels, we apply a mathematical correction based on the T1 dynamics of the blood pool. This step ensures that the resulting T1 values more accurately reflect the underlying tissue state rather than systemic blood variations.
Experimental Setup and Validation
The study utilized a massive dataset of 14,683 subjects from the UK Biobank, all scanned on 1.5T Siemens systems using the ShMOLLI sequence. The models were trained and validated using a subset of 1,600 manually segmented cases:
- Training: 800 subjects (including healthy controls and various cardiovascular conditions).
- Segmentation Validation: 100 subjects.
- QC Validation: 700 subjects.
After validation, the pipeline was deployed to perform inference on the remaining ~13,000 scans to establish population-level reference ranges.
Results and Clinical Insights
Performance Metrics
The automated pipeline demonstrated high fidelity to expert manual analysis. The segmentation achieved an average Dice metric of 0.84 for the LV myocardium, with a very high correlation (r=0.97) between automated and manual T1 measurements. The QC system proved exceptionally robust, achieving a balanced accuracy of 97.1%. Specifically, it caught 90.1% of erroneous outputs (sensitivity) while maintaining a specificity of 96.4%, ensuring that very few valid scans were unnecessarily rejected.
Population-Scale Findings
Out of the 14,683 exams processed, the QC system rejected only 156 scans (approximately 1%), demonstrating the high reliability of the underlying segmentation model. The resulting data allowed for the creation of the largest T1 reference database to date. Our analysis confirmed that native T1 values are significantly elevated in patients with Dilated Cardiomyopathy (DCM), Hypertrophic Cardiomyopathy (HCM), and acute myocarditis compared to healthy subjects. Furthermore, the inclusion of blood pool correction significantly improved the statistical separation between healthy and diseased cohorts.
Discussion and Limitations
This work demonstrates that Bayesian uncertainty is not merely a theoretical construct but a practical tool for clinical deployment. By quantifying "what the model doesn't know," we can build automated systems that are safe for large-scale population studies and clinical decision support.
However, some limitations remain. The current model was trained exclusively on UK Biobank data (Siemens 1.5T, ShMOLLI). Future work will focus on fine-tuning the framework for multi-vendor and multi-sequence compatibility to ensure generalizability across different hospital settings. Additionally, while native T1 is a powerful biomarker, future studies incorporating paired Late Gadolinium Enhancement (LGE) data will help further define the comparative diagnostic value of automated T1 mapping in detecting focal versus diffuse fibrosis.
Conclusion
The proposed pipeline enables the rapid, objective, and robust analysis of myocardial native T1 mapping at scale. By integrating automated QC, we provide a framework that can handle the complexities of real-world clinical data, paving the way for more precise and accessible cardiac tissue characterisation.
Summary updated