Attenuation Map Generation with Cross-Vendor and Cross-Tracer Transfer Learning for Cardiac SPECT

Previous studies have demonstrated the feasibility of generating attenuation maps from emission data using deep learning for cardiac SPECT. However, the deep-learning-based approaches were based on the assumption that the training and testing datasets were acquired using the same scanner and tracer. This would limit utility in clinical practice since it would be tedious and expensive to re-collect sufficient data each time for new tracers or scanners. In this work, we pre-trained a network with sufficient data acquired from a GE NM/CT 850 SPECT/CT scanner following the injection of 99mTc-tetrofosmin, then applied this network for attenuation map generation for Philips SPECT/CT scanner following the injection of 99mTc-sestamibi. Cross-tracer and cross-vendor transfer learning approaches were also investigated to further improve the performance. To optimize the transfer learning, input and output layers, input and output convolutional layers, and the whole network were fine-tuned respectively to search for the best modes. Three groups were compared: 1) trained with 120 GE data and tested with 30 Philips data without fine-tuning; 2) cross-validation with 10 Philips data for training and 20 Philips data for testing; 3) pre-trained with 120 GE data and fine-tuned by 10 Philips data, then tested with 20 Philips data. All three groups output accurate attenuation-corrected images. The averaged normalized mean squared error of the attenuation-corrected images in the transfer learning group compared to the ground-truth corrected with CT-derived attenuation maps was 1.07 ± 0.61%, better than 1.57 ± 0.63% and 1.65 ± 1.02% in the two groups without fine-tuning. This work demonstrated that it is feasible to apply a network trained by one scanner data to another scanner data for attenuation map generation. When possible, transfer learning with proper fine-tuning further improved the prediction accuracy.