Abstract 4139767: Automated IVUS Image Analysis for Cardiac Allograft Vasculopathy Surveillance in Heart Transplant Recipients

Background: Cardiac allograft vasculopathy (CAV) is a heterogenous and poorly understood process of immune and non-immune factors occurring within the coronary artery and a leading cause of graft failure in heart transplant recipients. Intravascular ultrasound (IVUS) is used as an adjunct to angiography to significantly increase the sensitivity for detection disease, however by current standards < 1% of IVUS images are graded for maximal intimal thickness (MIT) or scored for severity of CAV. The majority of IVUS frames remain unanalyzed. Aims: Demonstrate the application of pre-trained DeepIVUS machine learning model to annotate IVUS images obtained from transplant patients, establishing the accuracy of this tool in a new patient population. Methods: We reviewed 963 clinically annotated IVUS frames from heart transplant recipients, 706 were selected after screening for completeness. DeepIVUS was used to segment the frames and create predictions of the vessel lumen, plaque area and intimal thickness (IT) measurements around the entire vessel lumen. To compare with expert annotations, typically a single caliper measurement, the 90th percentile value of DeepIVUS-predicted IT was used. A confusion matrix was used to assess the model's ability to accurately classify predicted and true IT values, utilizing a threshold of 0.5 mm. Results: In 163/706 frames (23.1%), DeepIVUS labeling resulted in segments with jagged or non-physiologic shapes. These were automatically discarded using a quantitative ovalness criteria, Figure 1. Agreement between DeepIVUS predicted MIT and expert measured MIT in the remaining 543 frames yielded a strong Pearson correlation coefficient of 0.81 with an R2 of 0.65, Figure 2A. The linear fit was used to determine the DeepIVUS IT cutoff (0.554mm) which corresponded to clinical cutoff of 0.5mm. Using this value, DeepIVUS correctly classified 86.6% of frames (sensitivity 85.1%, specificity 87.9%, aF1 0.86), Figure 2B. Conclusion: We present a first use applcation of a pre-trained Deep IVUS image analysis in a transplant population demonstrating good agreement with expert clinical annotation. In current practice, a small fraction of images from each IVUS study are annotated. With an ovalness criteria providing quality control, DeepIVUS may be applied to non-annotated frames, allowing significantly more frames per IVUS study to be analyzed. Future studies are warranted to understand the prognostic capabilities of these automated annotations.