Fast AI Based Amyloid PET Reporting – can Nuclear Medicine Handle the Expected Workload in Alzheimers Diagnostics?

Ziel/Aim: With the approval of disease-modifying and in the future potentially curative therapies for Alzheimer’s, there is a looming threat of insufficient diagnostic capacities to assess treatment eligibility and monitor therapy success. Nuclear medicine’s PET modality is one of the gatekeepers. There is serious doubt whether we can actually cope with predicted demand. While currently commercially available whole-body PET/CTs can image a patient in less than one minute, is it feasible those images could be read and the report be written in the same amount of time? Since this task is expensive, time consuming and boring, we focused on finding a minimal viable AI tool that could perform the task of delineating pathologic from normal findings.

Methodik/Methods: We found that a 3-level deep learning binary classifier network with a sigmoid activation function at the last stage is sufficient to be trained within minutes and provide very high accuracy even for previously unseen PET images. For training, images of 20 Patients were used. They were imaged on a digital PET/CT using 18F-florbetabene as tracer with one bed of 1 minute and low-dose CT for attenuation correction. We trained with 196 amyloid-positive images and 100 amyloid-negative images for 100 epochs, found no signs of over-fitting even though accuracy was 100% on both test and training data while loss was 0. Validation with real world data of 5 patients not in the test or training data set showed 100% accuracy as well.

Ergebnisse/Results: The deep-learning network was capable of learning the task of delineating pathologic and normal findings within minutes on a standard PC and performing the classification task with 100% accuracy in less than 1 millisecond per case.

Schlussfolgerungen/Conclusions: Given the extremely fast training, the low cost to setup and maintain AI is very easy to deploy in clinical practice. Reports are finished almost instantly as the patient leaves the scanner. This doesn’t replace a trained professional, but can filter out clear cut cases that don’t need in-depth analysis. We can easily take our key role as gatekeeper for therapy of Alzheimer’s and other neurodegenerative diseases with a little help from AI and modern PET scanners. The AI filters out easy cases and frees time for those cases that require expertise.

Publikationsverlauf

Artikel online veröffentlicht:
12. März 2025