RT Journal Article
SR Electronic
T1 Creation of machine learning based classifiers for interpretation of I-123 Ioflupane images
JF Journal of Nuclear Medicine
JO J Nucl Med
FD Society of Nuclear Medicine
SP 1064
OP 1064
VO 62
IS supplement 1
A1 Nakajima, Kenichi
A1 Saito, Shintaro
A1 Chen, Zhuoqing
A1 Komatsu, Junji
A1 Inaki, Anri
A1 Watanabe, Satoru
A1 Kinuya, Seigo
YR 2021
UL http://jnm.snmjournals.org/content/62/supplement_1/1064.abstract
AB 1064Objectives: I-123 Ioflupane (FP-CIT, DATSCAN) has been an essential radiotracer for differentiating neurodegenerative diseases including Parkinson syndrome/diseases and Lewy body diseases. The abnormality of SPECT images has been commonly interpreted visually with semiquantitative uptake indices. To improve the recognition of abnormal patterns, we aimed to create a machine-learning method comparable to expert interpretation. Methods: Consecutive 108 patients (age, 67 ± 15 years; 44 males) were retrospectively selected. The transaxial DICOM images were preprocessed to select slices of the highest count on the striatum, and cropped and trimmed automatically for subsequent data processing. The images were then classified into 4 characteristic patterns, namely 1) the accumulation is high or low, 2) right and left uptake pattern is symmetric or asymmetric, 3) shape of the striatal uptake is comma-like (uptake in caudate and putamen) or dot-like (uptake of caudate), and 4) total impression is normal or abnormal. Various machine learning algorithms were applied including methods of decision tree, gradient boosted trees, logistic regression (LR), naïve Bayes (NB), nearest neighbors, neural network, random forest, and support vector machine (SVM). The datasets were trained and validated using 75% and 25% datasets, respectively, and the processing of random data sampling, training, and validation was repeated 10 times. Appropriate classifiers were determined for each image feature. Results: Machine learning methods were screened using measures of accuracy, Cohen Kappa, F1 score, area under the receiver operating characteristic curves (ROC-AUC), and candidate methods were selected. The image pattern of high/low showed the highest diagnostic accuracy with NB (ROC-AUC 0.93 ± 0.05 for high, 0.91 ± 0.08 for low) and LR (0.94 ± 0.05 for high and low) (p = ns). The symmetric/asymmetric patterns showed the highest diagnostic accuracy with nearest neighbors (0.75 ± 0.09 for symmetric, 0.75 ± 0.11 for asymmetric) and LR (0.78 ± 0.07) (p = ns). The image pattern of comma/dot showed the highest diagnostic accuracy with SVM (0.92 ± 0.05 for comma and dot) and LR (0.93 ± 0.06 for both) (p = ns). Lastly the final judgement of normal/abnormal was the highest with LR (ROC-AUC 0.93 ± .06). Although the NB method showed a higher accuracy (0.85 ± 0.06 by NB vs. 0.67 by LR), variation of ROC-AUC with NB was larger by 10 repeated measurements. Conclusions: For the image interpretation of DATSCAN, machine learning can be used as an effective pattern-recognition method with ROC-AUC of 0.92 - 0.94 using LR and/or SVM methods, and symmetricity could be diagnosed either with nearest neighbors or LR method. Machine learning can be a promising adjunctive tool as a DATSCAN image classifier when optimal training methods are applied.