Specific Binding Ratio Estimation of [¹²³I]-FP-CIT SPECT Using Frontal Projection Image and Machine Learning

Abstract

This study aimed to develop a new convolutional neural network (CNN) method for estimating the specific binding ratio (SBR) from only frontal projection images in single-photon emission-computed tomography using [123I]ioflupane. We created five datasets to train two CNNs, LeNet and AlexNet: (1) 128FOV used a 0° projection image without preprocessing, (2) 40FOV used 0° projection images cropped to 40 × 40 pixels centered on the striatum, (3) 40FOV training data doubled by data augmentation (40FOV_DA, left-right reversal only), (4) 40FOVhalf, and (5) 40FOV_DAhalf, split into left and right (20 × 40) images of 40FOV and 40FOV_DA to separately evaluate the left and right SBR. The accuracy of the SBR estimation was assessed using the mean absolute error, root mean squared error, correlation coefficient, and slope. The 128FOV dataset had significantly larger absolute errors compared to all other datasets (p < 0. 05). The best correlation coefficient between the SBRs using SPECT images and those estimated from frontal projection images alone was 0.87. Clinical use of the new CNN method in this study was feasible for estimating the SBR with a small error rate using only the frontal projection images collected in a short time.

Keywords

• dopamine transporter (DAT)
• single-photon emission-computed tomography (SPECT)
• imaging
• deep learning (DL)
• convolutional neural network
• specific binding ratio (SBR)