Employing transfer learning to optimize deep learning models for Alzheimer’s disease classification using two tau PET tracers

Abstract Background Deep learning models, particularly convolutional neural networks (CNNs), have shown promise in Alzheimer’s disease (AD) classification using tau PET data. However, limited sample sizes and unharmonized tau tracers present challenges to developing an agnostic tau tracer tool to predict AD using machine learning. Transfer learning, which leverages pre‐trained models for related tasks, may address these issues. Here we evaluate the effectiveness of transfer learning in optimizing 3D CNNs for AD classification with distinct cohorts and tau tracers. Method We used tau PET images from ADNI ([18F]Flortaucipir, n = 437) and TRIAD ([18F]MK‐6240, n = 423) cohorts, categorizing patients into CU (cognitively unimpaired) and CI (cognitively impaired). Standardized uptake value ratios (SUVR) were used for tau PET data. Separate 3D CNNs were trained for each tracer, with SUVR volumes as input and diagnosis as output. For transfer learning, we trained a model on [18F]Flortaucipir data with a reduced learning rate, using a pre‐trained model from [18F]MK‐6240. Models underwent 5‐fold cross‐validation, and metrics were computed as the average of validation metrics across folds. To avoid data leakage, images from the same subject were assigned to the same fold. Result The model trained on [18F]MK‐6240 tracer demonstrated higher classification performance than [18F]Flortaucipir (AUC = 0.84 vs 0.67; Figure 1. F1‐score = 79.77% vs 64.66%; Table 1). To enhance the classification performance of [18F]Flortaucipir model, we employed a transfer learning approach by leveraging the model pre‐trained with [18F]MK‐6240. With this approach, we observed a slight improvement in all classification metrics compared to the model trained solely on [18F]Flortaucipir data (AUC = 0.71 vs 0.67; Figure 1. Accuracy = 71.39% vs 67.72, F1‐score = 67.97% vs 64.66%; Table 1). Conclusion This finding highlights the value of transfer learning in optimizing deep learning models for Alzheimer’s disease classification, particularly when handling tau tracers with varying performance levels. Our results are consistent with previous on transfer learning’s effectiveness in this context. These preliminary findings indicate that applying this technique to larger datasets of tau tracers may further enhance model performance, potentially leading to the development of a tau tracer‐agnostic tool that overcomes the need of tracer harmonization for predicting dementia.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.