Abstract 4146737: AmyNet: a Novel Deep Learning Imaging Approach to Identify Incidental Myocardial Uptake Indicative of Cardiac Amyloidosis on Whole Body 99Tc-Bone Scintigraphy Imaging

Introduction: Whole-body scintigraphy (WBS) is frequently performed for oncological indications but also has high sensitivity for incidental detection of transthyretin amyloid cardiomyopathy (ATTR-CM), which may be overlooked. Goal: Develop an algorithm that combines chest computed tomography (cCT) and WBS to identify incidental ATTR-CM in a large healthcare system. Methods: We included 3 cohorts: 1) 77 patients from Brigham and Women’s Hospital with high prevalence of patients positive for ATTR-CM (BWHC), 2) 982 continuous cohort of patients imaged with WBS at PennMedicine for oncological indications from 2010-2020 (PennC), and 3) 5 patients with proven ATTR-CM uptake on WBS when imaged for oncologic indications from 2021-2024 (PennC+). Using a U-Net model trained with Dice score, the AmyNet algorithm automatically segmented cCTs with regions of interest (ROIs) around the entire heart (ventricles, atria, blood pool). Maximal intensity projections of cCTs in the coronal plane were registered to WBS using rigid transformations from the ANTs library. Multiple template reconstructions of only BWHC projections augmented registration due to limited field of view of cCT by providing an initial alignment to the thorax. Transformations were applied to heart ROIs to identify the location on WBS. A modified contralateral ratio was calculated by taking the lateral 1/3 and inferior 2/3 region to avoid sternum and atria and reflecting it across the vertical axis. A cutoff ratio of 1.15 was established as positive for ATTR-CM. Results: Machine-generated and expert volumes were comparable (1071vs1087mL; p=0.74) with a Dice similarity index of 0.92±0.07. AmyNet had a 94% accuracy with 100% sensitivity and 90% specificity in detecting ATTR-CM in BWHC with positive patients having higher contralateral ratios (1.57±0.49 vs 1.01±0.18, p<0.0001). Screening of PennC identified one patient (0.1%) with myocardial uptake, which was confirmed ATTR-CM upon evaluation for HFpEF. All 5 PennC+ patients were detected positive for uptake by AmyNet. Conclusion: We created the first algorithm that leverages cCT data and WBS to detect incidental cardiac uptake suggestive of ATTR-CM in patients underdoing scans for other non-cardiac indications.