Long Short Term Memory (LSTM) architecture based neural network encoder model for reducing noise in 1000fps High Speed Angiography image sequences

Diagnosis and image-guided endovascular treatment of vascular diseases such as strokes, and aneurysms can significantly benefit from 1000 fps High-Speed X-ray Angiography (HSAngio). This technology enables real-time visualization of flow and derivation of critical data, such as detailed velocity maps, while the patient is still on the procedure room table. Such information can assist interventionalists in assessing disease severity, planning treatments, and evaluating treatment prognosis, allowing for immediate adjustments to treatment devices for optimal outcomes. However, images acquired at such high frame rates suffer from substantial quantum mottle, limiting the visibility of low-contrast objects and accuracy of flow quantification. We propose a Long Short Term Memory (LSTM) architecture based neural network to mitigate noise in high-speed x-ray images. To train and evaluate the network, we simulated 1ms x-ray images (with and without quantum noise) of iodine contrast flow in three different patient-derived Internal Carotid Artery geometries using Computational Fluid Dynamics (CFD) modeling. From two geometries, we generated a dataset of 250,000 temporal sequences for network training over 300 epochs, while the third geometry was reserved for validation. The performance of our LSTM network was compared to a conventional filtering method combining spatial and temporal recursive schemes. The Mean Square Error (MSE) between the noiseless image and the image processed by the LSTM network demonstrated superior performance over the combined method for various regions of interest over the vessel in a 300-image sequence. In regions where pixel intensities changed due to varying contrast concentrations, the LSTM network consistently outperformed the combined method. Additionally, the temporal profiles indicated that the LSTM network preserved temporal information, whereas the combined method introduced a noticeable lag. The effectiveness of the LSTM network in reducing noise was also successfully demonstrated on actual images of contrast flow in a swine renal artery acquired at 1000 fps.