Objective: Our objective is to develop an AI-based approach to automatically rate FIP-CIT-SPECT data.

Background: FP-CIT-SPECT visualizes the integrity of nigro-striatal dopaminergic synapses in vivo. Although it is currently the gold-standard imaging method for the diagnosis of Parkinsonism, it is susceptible to inter-rater variability and quantitative threshold-values are not well established. Automated image processing by artificial intelligence (AI) holds the potential to drastically simplify and standardize FP-CIT-SPECT reading.

Method: 460 FP-CIT-SPECT datasets that had been acquired on a Picker Prism 3000 and reconstructed with an OSEM algorithm were retrospectively selected and expert ratings regarding presence of pathology were obtained and classified in three categories (healthy, pathologic, undetermined). Cranio-caudal maximum intensity projections (MIPs) were calculated to provide two-dimensional input images. A Google-Tensorflow® AI-environment was set up and the openly available pretrained “Inception-Net” for automated image classification was retrained with the FP-CIT MIPs after removal of the last training layer. Classification performance was cross-validated by a leave-n-out approach.

Results: Initial classification performance for automated detection of pathological scans was in a range from 65-75%. The application of other pre-trained image classification networks, classification of MIPs derived from spatially normalized data and alternative tree-classifier-driven approaches are currently being explored.

Conclusion: Although the classification performance of anatomically unprocessed data needs further improvement, we provide a proof of principle, that AI-based image processing in FP-CIT SPECT is possible. Improved image preprocessing and alternative approaches will likely have a significant impact on the classifications performance so that AI-based automated image rating of FP-CIT-SPECT might be supportive in a future clinical setting.

« Back to 2019 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/automated-interpretation-of-fp-cit-spect-images-based-on-artificial-intelligence/