Evaluation of BTXBrain-Dopamine, AI-powered automated quantification software for dopamine transporter PET

Introduction: Dopamine transporter (DAT) imaging is widely performed to evaluate the function of presynaptic dopaminergic neurons and is particularly useful in the differential diagnosis of Parkinson’s disease (PD) and other degenerative parkinsonism from essential tremor, vascular parkinsonism, and drug-induced parkinsonism. [¹⁸F]FP-CIT has several advantages over [¹²³I]FP-CIT, mainly due to the better spatial resolution and sensitivity of PET compared to SPECT [1]. Quantification of brain PET images based on spatial normalization (SN) techniques and predefined brain atlases allows for more objective and convenient assessment of regional PET activity than manual ROI drawing. However, direct spatial normalization of [¹⁸F]FP-CIT PET images onto the average template has not been recommended due to overestimation of striatal [¹⁸F]FP-CIT binding in PD patients [2]. Although an alternative SN method using early-phase [¹⁸F]FP-CIT PET image was proposed, the early-phase images are not always available [3]. Brightonix Imaging’s BTXBrain-Dopamine software (https://brtnx.com/en/product/product_ai_dopamine.php) allows automatic quantification of [¹⁸F]FP-CIT PET images using an artificial intelligent (AI)-based robust direct PET SN technology that does not need anatomical images. The purpose of this study is to evaluate the accuracy of BTXBrain-Dopamine software in the quantification of striatal [¹⁸F]FP-CIT binding.

Methods: BTXBrain-Dopamine’s SN engine consists of convolutional neural networks that generate deformation fields required for PET SN. Although the network model was trained using paired datasets of [¹⁸F]FP-CIT PET and 3D structural MRI, only the PET images are required as input in the inference stage. In addition, the inverse deformation from template space to individual space is also possible due to the cyclic training strategy. A total of 93 pairs of [¹⁸F]FP-CIT PET and 3D MRI datasets were used to evaluate the accuracy of striatal [¹⁸F]FP-CIT binding quantification as well as the striatal region segmentation. For the comparison, we used FIRST algorithm that segments striatal regions from 3D MRI, which were applied to the PET images in individual space.

Results: Striatal regions transformed from standard space using the inverse deformation fields generated by PET SNs matched well with regions determined by the segmentation of individual MRIs. Mean Dice coefficients between striatal regions determined by BTXBrain using only PET and FIRST using MRI were 0.83, 0.84, 0.78, and 0.78 for left putamen, right putamen, left caudate, and right caudate, respectively. The striatal counts determined by BTXBrain and MRI-based PET quantification using FIRST algorithm were highly correlated. R² and slope of linear regression between them ranged 0.97-0.99 and 1.00-1.03, respectively. The computation time for the SN and striatal binding ratio quantification was only 10-20 seconds in BTXBrain.

Conclusions: BTXBrain-Dopamine allowed for fast and automatic quantification of striatal binding ratio in [¹⁸F]FP-CIT PET images. The PET counts in putamen and caudate quantified by BTXBrain-Dopamine were virtually equivalent to those obtained by performing MRI segmentation and PET/MRI co-registration. Therefore, BTXBrain-Dopamine will be useful for evaluating presynaptic dopaminergic function in PD and other degenerative parkinsonism.

• Download figure
• Open in new tab
• Download powerpoint