Dense Optical Flow and U-Net Segmentation for Automated Circumferential Strain Mapping in Abdominal Aortic Aneurysms

Technical Summary: Dense Optical Flow and U-Net Segmentation for Automated Circumferential Strain Mapping in Abdominal Aortic Aneurysms

Problem Statement
Abdominal Aortic Aneurysm (AAA) rupture carries a high mortality rate, yet current clinical decision-making relies on maximum diameter (55 mm), which lacks sensitivity for predicting rupture in smaller aneurysms. Biomechanical indices, specifically circumferential wall strain, offer superior predictive value. However, deriving these indices from 2D B-mode ultrasound (US) is hindered by two critical bottlenecks:

1. Manual Segmentation: Current workflows require trained operators to manually delineate the AAA lumen boundary (5–15 minutes per sequence), preventing scalability and real-time application.
2. Sparse Tracking Limitations: Dominant motion tracking methods, such as the Sparse Demons (SD) algorithm, rely on feature points selected based on image gradients. This approach fails in lateral wall regions where low echogenicity (due to specular reflection geometry) results in poor point selection, degrading strain accuracy precisely where heterogeneity is most clinically relevant.

Methodology
The authors propose a fully automated, end-to-end pipeline that operates exclusively on open-source data and tooling. The workflow consists of five sequential stages:

1. Automated Segmentation (U-Net):

   • Architecture: A standard U-Net with four encoder/decoder levels, batch normalization, and skip connections.
   • Training Strategy: The model is pre-trained on the public CAMUS echocardiography dataset (500 patient sequences) to learn US speckle texture and boundary geometry. It is then fine-tuned on simulated AAA ultrasound sequences generated using the MUST Matlab toolbox.
   • Data: The simulated dataset includes 80 cine-loops with ground-truth displacement fields derived from patient-specific Finite Element Method (FEM) simulations.
   • Loss Function: A combination of Binary Cross-Entropy (BCE) and Soft Dice loss to address class imbalance and optimize overlap.

2. Dense Optical Flow Tracking:

   • Algorithm: The pipeline replaces sparse tracking with bilateral TV-L1 dense optical flow (implemented via OpenCV's DualTVL1OpticalFlow).
   • Mechanism: This variational method computes a displacement vector for every pixel, minimizing data fidelity (L1 norm) and Total Variation (TV) regularization. This ensures piecewise-smooth flow fields robust to speckle outliers and blood-pool noise.
   • Masking: The U-Net output defines a 6-mm annular Region of Interest (ROI) around the inner wall, restricting strain computation to the vessel wall and excluding blood motion.

3. Strain Computation:

   • Displacement fields are sampled on a regular grid within the ROI.
   • Circumferential strain is calculated using Radial Basis Function (RBF) interpolation (thin-plate splines) to derive Green's strain tensor components, avoiding the need for a regular mesh.
   • Cardiac cycle phases (end-diastole and peak-systole) are automatically detected by identifying extrema in the antero-posterior diameter curve derived from the segmentation masks.

Key Contributions

1. First Fully Automated Pipeline: To the authors' knowledge, this is the first end-to-end system for AAA wall strain mapping that eliminates manual intervention entirely.
2. Transfer Learning from Echocardiography: Demonstrates that a U-Net pre-trained on cardiac (CAMUS) data can effectively segment AAA walls when augmented with simulated AAA frames, achieving high Dice scores without requiring large annotated AAA datasets.
3. Dense vs. Sparse Tracking: Replaces the point-selection bias of Sparse Demons with dense optical flow, specifically addressing the tracking degradation in low-echogenicity lateral wall regions.
4. Open Science: All code, pre-trained weights, and synthetic dataset generation scripts are released publicly to ensure reproducibility.

Results
Evaluated on held-out simulated sequences, the proposed pipeline demonstrated significant improvements over baselines (Manual + Sparse Demons):

• Segmentation: The U-Net achieved a mean Dice coefficient of 0.908 ± 0.021 and an Intersection over Union (IoU) of 0.849 ± 0.028, significantly outperforming active contours (Dice 0.741) and CT-pretrained nnU-Net (Dice 0.837).
• Tracking Accuracy: Dense Optical Flow reduced the tracking Root-Mean-Squared Error (RMSE) to 0.087 ± 0.031 mm, a 39% reduction compared to the Sparse Demons baseline (0.142 mm). This improvement was most pronounced in the lateral wall regions.
• Strain Accuracy: The Mean Absolute Error (MAE) for circumferential strain decreased from 1.81% (baseline) to 0.93%.
• Computational Efficiency: The full automated pipeline reduced per-frame processing time from 180 ms (manual + SD) to **68 ms**, enabling near real-time operation (equivalent to 15–25 fps latency) on a single NVIDIA A100 GPU.

Significance and Claims
The paper claims that this pipeline addresses the two primary barriers to the clinical translation of US-based AAA biomechanics: operator dependency and tracking failure in low-echogenicity regions. By achieving sub-millimeter tracking accuracy and near real-time throughput using only open-source resources, the method facilitates reproducible research and lays the groundwork for prospective clinical validation. The authors note that while the 0.93% strain error is clinically meaningful (reducing relative error to <10%), the current evaluation is limited to synthetic data, and future work must validate the system on prospective clinical cohorts, particularly addressing 3D out-of-plane motion and intraluminal thrombus (ILT).