TY - JOUR T1 - Application of convolutional neural network to<sup>123</sup>I-MIBG SPECT imaging: automatic quantitation vs. manual measurements JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 1430 LP - 1430 VL - 61 IS - supplement 1 AU - Shintaro Saito AU - Kenichi Nakajima AU - Lars Edenbrandt AU - Olof Enqvist AU - Johannes Ulen AU - Seigo Kinuya Y1 - 2020/05/01 UR - http://jnm.snmjournals.org/content/61/supplement_1/1430.abstract N2 - 1430Introduction: 123I-metaiodobenzylguanidine (MIBG), analogue of noradrenaline scintigraphy has a long history of clinical use in Japan, and heart-to-mediastinum ratio and washout rate (WR) are common indicators of the sympathetic nervous activity of the heart. These are essential not only for the estimation of severity and prognosis of heart failure but also the diagnosis of movement disorders and dementia related to Lewy-body dementia and Parkinson disease. However, quantitative values are variable due to manual settings, which resulted in diagnostic uncertainty. Therefore, a convolution neural network (CNN)-based algorithm was developed. The diagnostic accuracy of this CNN approach was compared to to the conventional quantification method. Methods: A total of 51 patients (64% men, mean age 67 years) with heart and neurological diseases were studied with 123I-MIBG early and delayed planar and single-photon emission-computed tomography (SPECT), including chronic heart failure with coronary artery disease, arrhythmia, familial amyloid polyneuropathy, Lewy-body dementia and Parkinson disease. First, using both early and delayed SPECT images a CNN model was trained to semantically segment the liver and both lungs. Second, the CNN analyzed the early and delayed image sets simultaneously and aligned them based on first CNNs output. While some cardiac activities were faint or no accumulation in the heart, a second CNN was trained to output a heart volume of interest (VOI) with the same relative location and size for early and delayed images. The model was trained and evaluated using four-fold cross validation. Third, we calculated WR taking the total and average counts in the heart VOI from early and delayed images. WR was calculated directly from the measured count values without the use of background or reference volumes. Finally, we investigated correlation with CNN WR and initial manually set WR (ground truth (GT) in this study), versus planar WR with and without background (BG) correction. Results: The CNN correctly identified cardiac regions in patients with normal uptake and even with faint uptake, but it failed to analyze three patients; a patient with a giant liver cyst; a patient with right highly elevated diaphragm, and a patient with a leakage in an antecubital region during administration of 123I-MIBG. Among the 48 patients, there was good correlation between CNN WR and planar WR without BG correction (r2 = 0.693) showing comparable WR values; planar WR = 0.89 (CNN WR) + 0.06). In addition,CNN WR versus planar WR with BG correction and CNN WR versus GT WR also showed good correlations (r2 = 0.658 and 0.596, respectively). The CNN successfully analyzed images even under conditions of low accumulation in the heart. Conclusions: The CNN WR had a strong correlation with planar WR without BG correction in various clinical settings, including heart and neurological diseases. The CNN-based method is operator-independent and could improve diagnostic accuracy by minimizing the gray zone of diagnostic results, and reducing variation in findings among expert interpretations. ER -