Icon Legend

This session is not in your schedule.

This session is in your schedule. Click again to remove it.

Biomedical Imaging and Instrumentation

Biomedical Imaging and Instrumentation - Poster Session D

Poster G4 - Quantitative Evaluation of IVUS Images Segmentations Generated by a Pre-Trained Neural Network (DeepIVUS) Improves Agreement with Clinical Evaluation

Friday, October 25, 2024

3:30 PM - 4:30 PM EST

Location: Exhibit Hall E, F & G

Presenting Author(s)

Wenzhou Lyu, WL

Undergraduate Researcher
University of California, San Diego
San Diego, California, United States

Co-Author(s)

Antoinette S. Birs, MD

Advanced Heart Failure and Transplant Physician
University of California, San Diego, United States
LA

Lawrence Ang

Associate Clinical Professor, Medicine
University of California, San Diego, United States
EA

Eric Adler

Professor of Medicine
University of California, San Diego, United States
EM

Ehtisham Mahmud

Professor Of Clinical, Medicine
University of California, San Diego, United States

Last Author(s)

Francisco Contijoch, PhD

Associate Professor
UC San Diego, United States

Introduction: Cardiac allograft vasculopathy (CAV) is a leading cause of graft failure in heart transplant patients. Intravascular ultrasound (IVUS) significantly enhances disease detection when used with angiography. However, less than 1% of IVUS images are evaluated for maximal intimal thickness (MIT) or CAV severity, leaving the majority of IVUS frames unanalyzed. Automated scoring creates the opportunity to evaluate the entire IVUS dataset but requires high quality annotations throughout the vessel of interest. Specifically, vessel branching and variations in image quality can preclude accurate analysis of a frame. DeepIVUS is a pre-trained neural network designed for evaluation of patients evaluated for percutaneous interventions. Therefore, it is unclear whether segmentations generated by a pre-trained machine learning model (DeepIVUS) can be enhanced to improve agreement with clinical annotations and enable systematic CAV evaluation of the entire vessel.

Materials and

Methods: 963 IVUS frames obtained from heart transplant recipients which had corresponding clinical annotations of maximal intimal thickness (MIT) were analyzed. Image quality for all frames was graded (by author ASB) blinded to DeepIVUS processing results. DeepIVUS segmentations were generated for all frames to create predictions of the vessel lumen, and plaque area. We investigated whether two steps could ensure high quality segmentations and improve agreement with clinically-measured maximal intimal thickness (MIT).

First, the shape of the lumen and vessel were analyzed in terms of their ovalness. Ovalness was quantified using the Dice score to compare each contour with its corresponding fitted oval, determined by the OpenCV package. Frames with either lumen or plaque segmentations of insufficient ovalness were discarded. The optimal ovalness cutoff was defined as the value that kept the highest number of frames while maintaining a high Concordance Correlation Coefficient (CCC) score with clinical measures (Figure 1). Second, segmentations provide intimal thickness measurements around the entire circumference of the vessel lumen. We investigated which percentile maximized agreement with clinical assessment. (Figure 2).

Results, Conclusions, and Discussions:

Results: Clinical grading identified 257 frames as being of limited image quality. The use of the 90th percentile thickness maximizes agreement in images as graded to be of high clinical quality (Table 1). Application of the ovalness criteria of 0.9593 to DeepIVUS segmentations removed 216 of 963 frames. Agreement between clinical grading and ovalness was 78.5% (Table 2). The ovalness criteria increased CCC from 0.61 to 0.73 with 90th percentile thickness (See Table 1).

Discussion: This study highlights the potential of DeepIVUS segmentations alongside postprocessing steps to enhance IVUS image analysis. By implementing percentile thickness measurement and a quantitative ovalness criterion, we improved the accuracy of segmentations by DeepIVUS in a unique population, heart transplant recipients at risk for cardiac allograft vasculopathy (CAV). Automated post-processing enables a more comprehensive and systematic evaluation of the IVUS datasets and could enable analysis throughout the vessels of interest. This could yield new metrics of disease such as overall thickening amount or enable new phenotyping of patients.

Conclusion: The automated evaluation of IVUS images using the DeepIVUS model significantly improves the agreement with clinical annotations. By implementing a quantitative ovalness criterion and percentile-based intimal thickness measurement, we demonstrated enhanced accuracy and reliability of IVUS image analysis, enabling detailed phenotyping of coronary allograft vasculopathy.

Acknowledgements (Optional):