AI Tackles Respiratory Motion in Radiotherapy with Mixed Results

Respiratory motion is a thorn in the side of radiotherapy accuracy for thoraco-abdominal tumors. With the latency of treatment systems introducing uncertainties in targeting, it's key to forecast frame motion accurately. This is where AI steps in, aiming to compensate for these delays. But does it succeed, or are we just slapping a model on a GPU rental?

RNNs and Transformers: A Mixed Bag

Researchers have turned to recurrent neural networks (RNNs) and transformers, two AI powerhouses, to tackle this issue. Why these? RNNs adapt to changing respiratory patterns with on-the-fly parameter updates, while transformers capture long-term dependencies in time-series data. That's the theory, at least.

The experiments involved 12 sagittal thoracic and upper-abdominal cine-MRI sequences from ETH Zürich and OvGU. The OvGU data set, known for higher motion variability and noise, presented a tough challenge. Principal component analysis (PCA) was employed to break down the Lucas-Kanade optical-flow field into static deformation modes and time-dependent weights.

Benchmarking the Methods

Several methods were tested for forecasting these weights. Linear filters, transformer encoders, and RNNs trained with advanced learning algorithms like real-time recurrent learning (RTRL) and sparse one-step approximation (SnAp-1) were on the table. The goal? Warp the reference frame to generate future images with minimal error.

Linear regression shined at short horizons, achieving a 1.3mm geometrical error at a 0.32-second forecast with the ETH Zürich dataset. However, for medium-to-long horizons, RTRL and SnAp-1 took the lead, keeping errors below 1.4mm and 2.8mm for ETH Zürich and OvGU sequences, respectively. The sequence-specific transformer held its own for low-to-medium horizons but struggled with data scarcity and domain shifts.

Too Soon to Celebrate?

So, should we be popping champagne corks? Hardly. While predicted frames visually matched the ground truth, notable errors sprang up near the diaphragm at end-inspiration and regions with out-of-plane motion. If the AI can hold a wallet, who writes the risk model on these forecasts?

Here's the stark reality: despite promising results, transformers and RNNs face an uphill battle against data limitations and the notorious variability of respiratory motion. Until these obstacles are addressed, the intersection is real, but ninety percent of projects might remain in the shadows.