Convolutional LSTM model for cine image prediction of abdominal motion

Area of Science:

• Medical Physics
• Radiotherapy Technology
• Artificial Intelligence in Medicine

Background:

• Latency in magnetic resonance linear accelerator (MR-Linac) systems hinders accurate target coverage in real-time radiotherapy.
• Precise motion prediction of abdominal organs is crucial for mitigating this challenge.

Purpose of the Study:

• To develop and evaluate a convolutional long short-term memory (convLSTM) model for enhanced motion prediction of abdominal organs.
• To improve accuracy in gated real-time radiotherapy by addressing MR-Linac system latency.

Main Methods:

• A convLSTM model with a sequence-to-one architecture was utilized, processing six input frames to predict one output frame.
• 2D cine magnetic resonance (cine-MR) imaging data from 17 patients were used, focusing on stomach, liver, pancreas, and kidney.
• Structural Similarity Index Measure (SSIM) was employed as the loss function, with performance evaluated using SSIM, NMSE, NCC, and PSNR.

Main Results:

• The convLSTM model achieved average SSIM values of 0.54 (stomach), 0.64 (liver), 0.77 (kidney), and 0.66 (pancreas) for predicting motion four frames ahead.
• Transfer learning with fine-tuning further improved SSIM values to 0.69 for the liver and 0.78 for the kidney.
• The model demonstrated accurate motion prediction for multiple abdominal organs using an Elekta Unity MR-Linac equivalent sequence.

Conclusions:

• The developed convLSTM model effectively predicts abdominal organ motion using cine-MR imaging.
• This predictive capability is vital for reducing latency in MR-Linac systems.
• The advancement holds potential for improving the precision and efficacy of real-time radiotherapy for abdominal cancers.