PT  - JOURNAL ARTICLE
AU  - Erika Minoshima
AU  - Taylor Wissler
AU  - Donna Cross
AU  - Satoshi Minoshima
TI  - The Development of Composite Artificial Intelligence (AI) without Data Sharing to Improve Classification of Amyloid PET in Alzheimer’s Disease (AD)
DP  - 2019 May 01
TA  - Journal of Nuclear Medicine
PG  - 1214--1214
VI  - 60
IP  - supplement 1
4099  - http://jnm.snmjournals.org/content/60/supplement_1/1214.short
4100  - http://jnm.snmjournals.org/content/60/supplement_1/1214.full
SO  - J Nucl Med2019 May 01; 60
AB  - 1214Objectives: The training of AI systems typically requires a large amount of aggregated data (‘big data’). It is challenging for researchers applying AI in molecular imaging since the amount of available data is often limited at individual institutions. This study explores the feasibility of a composite AI system to improve amyloid PET classification by aggregating multiple AI systems trained individually with small data sets instead of pooling data. Materials and Methods: 3D-SSP extracted regional values from [F-18]florbetapir PET data (ADNI, 165 normals, age 75±7 yrs, 78 female and 150 AD patients, age 75±8 yrs, 63 female, normalized to the cerebellum) were used to train 4 independent AI cores (multi-layer perceptron, 5 layers, logistic activation function). Simulating a multi-institutional setting, each AI system was trained with 10 different case mixes (combinations of AD vs normals with 150, 145, and 135 case mixes) randomly chosen from the master database. Individual networks (weights and activation thresholds) were then extracted and aggregated (non-weighted and weighted arithmetic means) across the systems to form the composite AI system. The classification accuracy of the individual AI systems vs the composite AI system was assessed using the common test data set. The categorization accuracy from the composite AI was compared to that from conventional SUVr analysis. Results: The classification accuracies of individual AI cores that were trained with the limited data sets varied substantially depending on the case mix of the training data sets (mean +/- SD, 74% +/- 10%; max 87%; and min 60%). The system trained with a relatively larger number of training cases (295 vs 285) generally resulted in better accuracy (79% vs 67%, respectively). In contrast, the composite AI system consistently outperformed individual AI cores (mean +/- SD, 87% +/- 2% vs 74% +/- 10%, respectively), resulting in an average of 19% improvement in accuracy, with the largest improvement being 43% as compared to the individual AI core performance. The composite AI system performed equally to the classification using the conventional SUVr (87% by the composite AI system vs 88% by SUVr using a threshold of 1.35). Conclusions: This study demonstrates that a composite AI system can be constructed based on multiple neural networks from individually trained AI cores, each using a small data set. The composite AI system outperforms individual AI cores in the classification of amyloid PET for Alzheimer’s disease. This approach can be applied to AI development across multiple institutions without necessitating data sharing or pooling to create a large training data set. Further investigation is warranted to test various algorithmic configurations of the composite AI system.