Improving Generalizability of PET DL Algorithms: List-Mode Reconstructions Improve DOTATATE PET Hepatic Lesion...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Deep learning (DL) for DOTATATE PET lesion detection requires large, annotated datasets, which are scarce due to low incidence of gastroenteropancreatic neuroendocrine tumors (GEP-NETs) and high annotation costs.
• DL models often lack generalizability, performing poorly on data from different PET/CT scanners or protocols, necessitating larger, more diverse training sets.

Purpose of the Study:

• To investigate the feasibility of enhancing DL algorithm performance by aligning background noise characteristics between training and out-of-domain testing datasets.
• To evaluate the impact of list-mode reconstructions on improving DL model generalizability for GEP-NET detection.

Main Methods:

• Acquired ⁶⁸Ga-DOTATATE PET/CT datasets from a modern digital scanner (Scanner1) and an older analog scanner (Scanner2).
• Reconstructed Scanner1 data using standard parameters and list-mode reconstructions at varying durations (2-5 min). Scanner2 data used standard iterative reconstruction.
• Trained DL networks on data from each scanner and tested Network1 (Scanner1) on out-of-domain Scanner2 data, evaluating performance with varying training data fractions.

Main Results:

• List-mode reconstructed data from Scanner1 (2-min) showed noise levels most similar to Scanner2 data, yielding the best DL performance (F1 = 0.713), comparable to in-domain training (F1 = 0.755).
• Increasing training data from 25% to 100% significantly improved DL model performance (F1 from 0.478 to 0.713; p < 0.001).
• List-mode reconstructions enable modern PET data to mimic noise properties of older scanners, improving DL generalizability.

Conclusions:

• Reconstructing modern PET/CT data with list-mode protocols can effectively match the noise characteristics of older scanners, enhancing DL algorithm performance.
• Utilizing existing annotated data with list-mode reconstructions reduces dataset generation costs and effort, significantly improving DL model generalizability for GEP-NET detection.