TY - JOUR T1 - Task-Based Assessment of 3-Dimensional Striatal Segmentation of MRI for PET Quantification using Multi-Task Learning JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 503 LP - 503 VL - 61 IS - supplement 1 AU - Mario Serrano-Sosa AU - Jared Van Snellenberg AU - Jiayan Meng AU - Mark Slifstein AU - Chuan Huang Y1 - 2020/05/01 UR - http://jnm.snmjournals.org/content/61/supplement_1/503.abstract N2 - 503Objectives: Dysfunction in the striatum has been studied in neuropsychiatric diseases such as schizophrenia and Huntington’s disease1-4, where functional imaging relies on proper segmentation for analyses. Manual segmentation of the striatum performed by experts is time-consuming and may show poor interrater reliability. Therefore, automated deep learning techniques have been used to segmented the striatum5; however, the performance was mainly evaluated using Dice index with no task-based analysis to determine the reliability of their segmentations for research use. More importantly, no deep learning technique for subregion segmentation has been reported. Multi-Task Learning (MTL) is a machine learning technique that allows representations between related tasks to be shared to generalize the model and predict the original task, and many more, with better accuracy6. In this study we propose to utilize MTL to segment subregions of the striatum consisting of pre-commissural putamen (prePU), pre-commissural caudate (preCA), post-commissural putamen (postPU), post-commissural caudate (postCA), and ventral striatum (VST), and perform task-based assessment of the segmentation performance. Methods: Striatal and extastriatal regions of interest (ROI’s) were manually drawn on individual T1-weighted scans using a method previously validated7,8. Manually drawing the striatum and its various subregions is a time-consuming task, with the task taking 1.5 days on average per scan. The 3D T1-weighted image was used as input for the MTL network and six output tasks were simultaneously trained; where five of the six were ROI’s masks and the last was the masked background. 87 datasets were split into 83 training, 4 validation. An independent dataset of 19 volunteers imaged with [11C]raclopride PET/MRI were used for the testing set. As shown in Figure 1, this MTL network consisted of a 3D uNet architecture. This model was trained to minimize the sparse softmax cross entropy between output and ground truth. Dice Similarity Coefficients (DSC) were utilized to initially evaluate the performance of the MTL network output. Afterwards, network segmentations were used for PET quantification. 60 min of dynamic emission data were acquired on an Siemens mCT scanner, following a bolus injection of 349 +/- 109 MBq of [11C]raclopride. Data were reconstructed by FBP with CT used for attenuation correction. Binding Potential (BPND) was derived in each ROI using simplified reference tissue model (SRTM)9 with cerebellum as reference tissue. Linear Regression was performed with BPND obtained from manual and CNN-generated ROIs. R2 and fractional differences were reported. Results: Table 1 shows mean DSC across the independent testing set with 0.9011, 0.9127, 0.9034, 0.8651, 0.8225, for prePU, preCA, postPU, postCA and VST, respectively; wherein VST showed lowest DSC since it is relatively small and, more importantly, has the most irregular shape out of all ROIs. Figure 2 shows an example of both manual and CNN-generated segmentations. PET quantitative analysis showed R2 between BPND using manual and CNN-generated ROI’s was ≥ 0.97 in all subdivisions of the caudate and putamen and 0.91 in VST. Figure 3 shows the percent differences of BPND between the manual and CNN-generated ROI’s. Conclusions: MTL based 3D Striatal Segmentation of MRI proposed here has been shown to provide fast and accurate segmentation for the purpose of PET quantification of the subregions of striatum that avoid the cumbersome task of hand drawing ROI’s. View this table:Table 1: Mean Dice scores (DSC) with standard deviations for all sub-region ROI's ER -