Deep learning pipeline for OCT plaque segmentation, enabling precise arterial risk assessment in cardiovascular imaging.

RayTrace Segment
Ray-casting segmentation for convex anatomical structures in MRI
Highlights
Clinicians and researchers needed a fast, reliable method to segment anatomical structures such as brain tumors and cardiac ventricles from MRI scans for diagnostic support and surgical planning.
Develop a ray-casting-based segmentation algorithm optimized for convex anatomical regions, using ensemble learning to enhance boundary detection under class imbalance.
Engineered a pipeline that converts 2D MRI data into 1D ray profiles, applied classifiers like AdaBoost.M2 and SMOTEBoost, and used spline interpolation to construct smooth anatomical masks.
Delivered accurate segmentations with up to 91.8% Dice similarity and millisecond-level runtime, demonstrating strong performance across two clinical datasets.
Core Team
Overview
This project addresses the challenge of segmenting anatomical structures in MRI images with high accuracy and low computational cost. Leveraging a ray-casting technique, the algorithm emits radial beams from a central point to detect boundary points of target regions like the heart's left ventricle and brain tumors.
To overcome data imbalance, a common issue in pixel-based classification, the approach incorporates ensemble learning methods like AdaBoost.M2, SMOTEBoost, and UnderBagging. The algorithm was tested on two major datasets: a cardiac MRI dataset from the University of York and a brain tumor MRI dataset from Southern Medical University. The results demonstrate impressive segmentation accuracy (up to 89.5% Dice coefficient) and fast execution times (as low as 4.1 ms), making it suitable for real-time or clinical scenarios.
Data
This project utilized two publicly available, clinically annotated MRI datasets to evaluate the segmentation algorithm's performance on cardiac and brain anatomical structures.
- •Cardiac MRI Dataset: Comprising 7,980 short-axis cardiac MRI images from 33 patients (York University). Each sequence includes 20 frames with 8–15 slices per patient at 256×256 pixels resolution. Manual segmentations of the left ventricle's endocardium and epicardium were provided, each described by 32-point contours.
- •Brain Tumor MRI Dataset: Contains 3,064 T1-weighted contrast-enhanced MRI slices from 233 patients (Southern Medical University), categorized into meningioma (708 slices), glioma (1,426 slices), and pituitary tumor (930 slices). Each image has a resolution of 512×512 pixels with manual tumor border delineations by experienced radiologists.
Methods
The segmentation algorithm follows a structured pipeline comprising five key stages: preprocessing, ray emission, data gathering, classifier training, and post-processing.
- Preprocessing: Histogram-based contrast enhancement using bimodal Gaussian filtering, intensity normalization to reduce variability, and resampling to ensure uniform resolution across datasets.
- Ray Emission and Data Gathering: From a central point within the region of interest, rays are emitted at uniform angular intervals (∆φ), sampling intensity profiles along their paths. The number of rays (e.g., 8, 16, 32, 64) determines the angular resolution (Figure 1).
- Classifier Training: The extracted patches exhibit class imbalance, with boundary points being underrepresented. Ensemble learning techniques such as AdaBoost.M2, RUSBoost, UnderBagging, SMOTEBagging, and SMOTEBoost were employed to improve boundary detection accuracy.
- Post-Processing: Detected boundary points are interpolated using cubic splines to generate smooth, closed contours representing the segmented anatomical structures.
Results
The algorithm produced high-quality segmentation results with low computational latency across both cardiac and brain tumor datasets.
- Heart Segmentation: Dice Score up to 91.8% (mean 84.7–85.0% across ray configurations). Optimal trade-off at ∆φ = π/8 with speed of 4.1–20.2 ms per slice (Figure 2).
- Brain Tumor Segmentation: Dice Score up to 89.5% (mean 82.2–83.0%). Consistent performance across tumor types with speed of 5.1–16.0 ms per slice (Figure 3).
- Best Classifiers: AdaBoost.M2 and SMOTEBoost offered the best trade-off between recall and precision across classifiers.
Conclusion
This project introduced a ray-casting segmentation algorithm that efficiently and accurately detects anatomical structures in MRI images. By reducing the task to one-dimensional boundary detection, the method achieved high Dice scores (up to 91.8% for the heart and 89.5% for brain tumors) with millisecond-level processing times.
The approach proved especially effective for convex shapes and benefited from ensemble classifiers like AdaBoost.M2 and SMOTEBoost to handle class imbalance. With a balanced configuration at ∆φ = π/8, it offers a strong trade-off between speed and accuracy. Future work could extend this method to non-convex structures and 3D imaging, further enhancing its clinical relevance.
Keep Exploring
Synthetic video with exact depth, alpha masks, optical flow and layered bokeh for training depth-aware video bokeh models, with a public dataset and a web demo.




