External evaluation of an automated quantification software for dopamine transporter PET without MRI

Introduction: One of the neurodegenerative diseases, Parkinson's disease, is often not identified until they have advanced considerably. One commonly used method for early detection of Parkinson's is [¹⁸F]FP-CIT PET imaging, which can reveal dopamine transporters in the striatum. However, analyzing [¹⁸F]FP-CIT scans can be challenging, as the striatal region is small in comparison to the entire brain, and detailed 3D structural MRI scans and manual segmentation are necessary for accurate quantification. Additionally, changes in uptake patterns as the disease progresses further complicate the diagnosis. To overcome the technical difficulties, we developed an automated quantification software utilizing deep learning-based spatial normalization, which does not require a detailed 3D T1 MRI. In this study, we provide the external validation results with different center data from the training set.

Methods: In this study, we used a pre-trained neural network and fine-tuned it using a dataset of 231 [¹⁸F]FP-CIT PET scans. The initial pre-training was performed using approximately 1000 amyloid PET images, obtained from the Korea National Information Society Agency (NIA) public database. The [¹⁸F]FP-CIT images used for fine-tuning were acquired from Seoul National University Hospital. The fine-tuned neural network was able to generate spatially normalized [¹⁸F]FP-CIT images in the Montreal Neurological Institute (MNI) space without the need for detailed structural MRI scans. To evaluate the performance of the quantification method, we used a dataset of 135 [¹⁸F]FP-CIT images that were paired with 3D T1 MRI scans. These scans were obtained from SMG-SNU Boramae Medical Center. In order to quantify the striatal uptakes, we used the FIRST algorithm to process the paired T1 MRI images and calculate the regional uptakes in individual spaces. These values were considered as the "ground truth" for our analysis. The FIRST algorithm was also used to process the MNI T1 template image, and the normalized images were overlapped with this segmentation to quantify the regional striatal uptake in the normalized space. We used linear regression to compare the ground truth values with the results obtained using the BTXBrain-DAT normalization method for the caudate nucleus and putamen.

Results: After the fine-tuning, the neural network was able to generate spatially normalized [¹⁸F]FP-CIT images in the Montreal Neurological Institute (MNI) space without the need for detailed structural MRI scans. The slopes of linear regression for left caudate, left putamen, right caudate and right putamen were close to the perfect linearity and all values were in the range from 0.95 to 1 (0.977, 0.983, 0.996, and 0.983). It also showed a strong relationship between the ground truth and measured data, where all r-squared values were higher than 0.95. (0.966, 0.989, 0.971 and 0.986).

Conclusions: In this study, we evaluated BTXBrain-DAT, an automated quantification software utilizing deep learning-based spatial normalization, with the external dataset. The results of our analysis showed that the proposed method had a strong relationship between the ground truth and the measured data, with R-squared values higher than 0.95 for all regions of interest, indicating that our automated method is a reliable and robust tool for quantifying [¹⁸F]FP-CIT PET images for Parkinson's disease diagnosis. This method can be useful for the early detection and follow-up of Parkinson's disease which is important for the treatment of patients. In the future, we will evaluate our software using more datasets from various centers and various demographics.

• Download figure
• Open in new tab
• Download powerpoint