PND78 AUTOMATED ASSESSMENT OF THE SUBSTANTIA NIGRA ON SUSCEPTIBILITY MAP-WEIGHTED IMAGING USING DEEP CONVOLUTIONAL NEURAL NETWORKS FOR DIAGNOSIS OF IDIOPATHIC PARKINSONâ⑬ⓢS DISEASE

Abstract

Despite the utility for determining nigrostriatal degeneration, lack of objective interpretation of nigrosome 1 susceptibility map-weighted imaging (SMWI) limits its generalized use. The purpose of this study is to implement and evaluate a deep learning (DL)-based approach for interpreting nigrosome 1 SMWI. This retrospective study enrolled 427 subjects (267 patients with idiopathic Parkinson’s disease (IPD) and 160 control subjects (patients with drug-induced parkinsonism and healthy subjects) at our institute, and additionally enrolled 25 IPD patients and 31 control subjects at other two institutes on approval of the local institutional review boards. All patients underwent both SMWI and dopamine transporter (DAT) imaging, while healthy subjects only underwent SMWI. DAT imaging served as reference standard for the presence or absence of abnormality on SMWI. Diagnostic performance was compared between a deep learning-based diagnostic algorithm and the visual assessment by an experienced neuroradiologist for both internal and external datasets by using “DTCompair” package of R. Per-subtantia nigra (SN) diagnostic sensitivity and specificity for determining abnormality in SN by an experienced neuroradiologist were 96.8% and 92.3%, and those by the developed algorithm were 97.7% and 89.7%, respectively (internal validation); 89.1% and 95.5%, and 93.5% and 92.4%, respectively (external validation). Per-participant diagnostic sensitivity and specificity for determining abnormality in SN by an experienced neuroradiologist were 99.1% and 93.3%, and those by the developed algorithm were 97.4% and 93.3%, respectively (internal validation); 96.0% and 93.6%, and 96.0% and 96.8%, respectively (external validation). Both per-SN and per-participant analyses showed differences of 90% lower confidence interval and 90% upper confidence interval within -0.1 – 0.1 for both sensitivity and specificity, indicative of comparable diagnostic performance. Our deep learning-based algorithm for determining the presence or absence of abnormality in SN on SMWI showed the diagnostic performance comparable to that by an experienced neuroradiologist.