Frame forecasting in cine MRI using the PCA respiratory motion model: comparing recurrent neural networks trained online and transformers

TL;DR

This study compares recurrent neural networks trained online and transformer models for predicting respiratory motion in cine MRI, aiming to improve radiotherapy targeting by compensating for treatment delays.

Contribution

It introduces a PCA-based motion model combined with various forecasting algorithms, including online-trained RNNs and transformers, for improved frame prediction in thoraco-abdominal cine MRI.

Findings

RTRL and SnAp-1 outperform other methods at medium-to-long horizons.

Linear regression is most accurate at short prediction horizons.

Transformers are competitive at low-to-medium horizons but limited by data scarcity.

Abstract

Respiratory motion complicates accurate irradiation of thoraco-abdominal tumors during radiotherapy, as treatment-system latency entails target-location uncertainties. This work addresses frame forecasting in chest and liver cine MRI to compensate for such delays. We investigate RNNs trained with online learning algorithms, enabling adaptation to changing respiratory patterns via on-the-fly parameter updates, and transformers, increasingly common in time-series forecasting for their ability to capture long-term dependencies. Experiments used 12 sagittal thoracic and upper-abdominal cine-MRI sequences from ETH Z\"urich and OvGU; the OvGU data exhibited higher motion variability, noise, and lower contrast. PCA decomposes the Lucas-Kanade optical-flow field into static deformation modes and low-dimensional, time-dependent weights. We compare various methods for forecasting these weights: linear filters, population and sequence-specific transformer encoders, and RNNs trained with real-time recurrent learning (RTRL), unbiased online recurrent optimization, decoupled neural interfaces, and sparse one-step approximation (SnAp-1). Predicted displacements were used to warp the reference frame and generate future images. Prediction accuracy decreased with the horizon h. Linear regression performed best at short horizons (1.3mm geometrical error at h=0.32s, ETH Z\"urich dataset), while RTRL and SnAp-1 outperformed the other algorithms at medium-to-long horizons, with geometrical errors below 1.4mm and 2.8mm on the sequences from ETH Z\"urich and OvGU, respectively. The sequence-specific transformer was competitive for low-to-medium horizons, but transformers remained overall limited by data scarcity and domain shift between datasets. Predicted frames visually resembled the ground truth, with notable errors occurring near the diaphragm at end-inspiration and regions affected by out-of-plane motion.