Abstract
The large number of available MRI sequences means patients cannot realistically undergo them all, so the range of sequences to be acquired during a scan are protocolled based on clinical details. Adapting this to unexpected findings identified early on in the scan requires experience and vigilance. We investigated whether deep learning of the images acquired in the first few minutes of a scan could provide an automated early alert of abnormal features. Anatomy sequences from 375 CMR scans were used as a training set. From these, we annotated 1500 individual slices and used these to train a convolutional neural network to perform automatic segmentation of the cardiac chambers, great vessels and any pleural effusions. 200 scans were used as a testing set. The system then assembled a 3D model of the thorax from which it made clinical measurements to identify important abnormalities. The system was successful in segmenting the anatomy slices (Dice 0.910) and identified multiple features which may guide further image acquisition. Diagnostic accuracy was 90.5% and 85.5% for left and right ventricular dilatation, 85% for left ventricular hypertrophy and 94.4% for ascending aorta dilatation. The area under ROC curve for diagnosing pleural effusions was 0.91. We present proof-of-concept that a neural network can segment and derive accurate clinical measurements from a 3D model of the thorax made from transaxial anatomy images acquired in the first few minutes of a scan. This early information could lead to dynamic adaptive scanning protocols, and by focusing scanner time appropriately and prioritizing cases for supervision and early reporting, improve patient experience and efficiency.
Introduction
Cardiovascular Magnetic Resonance (CMR) imaging is a highly versatile diagnostic tool, yet its complexity necessitates pre-defined scanning protocols based on a patient's known history. A significant challenge arises when unexpected abnormalities—such as aortic dilatation or hidden pleural effusions—are visible in the initial planning (scout) images but go unnoticed until after the scan is complete. Currently, adapting a protocol in real-time requires the constant presence of an experienced clinician or radiographer. This study investigates whether a deep learning system can analyze these early anatomy images to provide an automated "early warning," potentially enabling scanners to dynamically adapt their protocols in real-time.
Methods
Study Design and Dataset
The researchers utilized a dataset of 575 sequential CMR scans from two hospitals, involving hardware from two major manufacturers (Siemens and Philips). The data was partitioned into a training/validation set of 375 scans and a separate testing set of 200 scans. To ensure real-world robustness, the dataset included scans with common artifacts, such as those resulting from atrial fibrillation or poor breath-holding.
Technical Pipeline
The developed pipeline transforms raw 2D anatomy slices into a comprehensive 3D clinical assessment through three primary steps:
- Segmentation: An adapted HRNet (High-Resolution Network) architecture was trained to identify nine anatomical classes, including cardiac chambers, the aorta, the pulmonary artery, and pleural effusions.
- 3D Reconstruction: Segmented 2D images are assembled into a 3D thoracic model using trilinear interpolation.
- Clinical Quantification: From this 3D model, the system calculates physiological metrics, such as ventricular volumes and vessel diameters.
To account for inter-vendor variability, a linear regression model was employed to calibrate AI measurements against final clinical reports, ensuring consistency across different hardware manufacturers.
Results
The system demonstrated high accuracy across the 200-scan test set. The mean Dice coefficient for segmentation was 0.910, indicating strong agreement with human expert annotations.
Key diagnostic performances included:
- Aorta Dilatation: 94.4% accuracy with 100% sensitivity.
- LV Dilatation: 90.5% accuracy.
- LV Hypertrophy: 85% accuracy.
- Pleural Effusions: Area under the ROC curve (AUC) of 0.91.
- Disease Detection: The model correctly identified 100% of Hypertrophic Cardiomyopathy (HCM) cases and 90% of Dilated Cardiomyopathy (DCM) cases.
Discussion
This proof-of-concept study demonstrates that "scout" or anatomy images, traditionally used only for planning, contain sufficient diagnostic signal for high-accuracy automated analysis. By achieving performance comparable to the inherent variability of human clinicians, this AI system offers a pathway toward real-time protocol optimization. For instance, early detection of a high mass-to-volume ratio could automatically trigger specialized sequences, such as T1-mapping for amyloidosis, before the patient leaves the scanner.
Limitations and Future Work
While promising, the study is limited by the relatively low resolution of anatomy slices compared to gold-standard cine sequences. Future work must focus on technical integration with MRI scanner software to enable real-time hardware communication and validation across a broader range of pathologies and clinical centers.
Summary updated