Prediction of Multileaf Collimator Leaf Positions Using a Hybrid BKA– Neural Network Model for Radiotherapy Delivery Accuracy

Radiation therapy is a powerful tool for treating cancer, using high-energy beams to destroy malignant cells while sparing healthy tissue. To make this treatment precise, doctors use a device called a multileaf collimator, which acts like a dynamic shutter made of many thin metal leaves. These leaves move in and out to shape the beam, carving out a custom path that matches the tumor's shape exactly. The goal is to deliver a lethal dose to the cancer while leaving the surrounding organs untouched. However, the machinery is not perfect. Just as a mechanical shutter might stick or lag slightly, these metal leaves can sometimes move to a position that is a fraction of a millimeter off from where the computer planned them to be. In the world of radiation, even a tiny error of one millimeter can change the dose the patient receives, potentially reducing the effectiveness of the treatment or harming healthy tissue. Because of this, ensuring the leaves are in the exact right place at the exact right moment is a critical challenge for medical physicists.

Researchers have long tried to predict these tiny movements before they happen, hoping to correct them in advance. They have tested various computer methods, including standard neural networks, which are systems modeled after the human brain's ability to learn from patterns. While these systems can learn, they often struggle to be both fast and accurate, especially when trying to predict the position of every single leaf in a complex machine. A new study by Jun Lv and colleagues introduces a different approach that combines a standard neural network with a nature-inspired search method called the Black-winged Kite Algorithm. This algorithm mimics the hunting behavior of a specific bird of prey, using its efficient search patterns to find the best possible starting settings for the computer model. By using this method, the researchers aimed to create a system that could predict the exact location of the leaves with far greater accuracy and speed than previous methods.

The team tested their new model using data from fifty real-world cancer treatment plans, covering tumors in the head and neck, chest, and abdomen. They fed the computer thirteen different pieces of information from the machine's logs, such as the angle of the radiation source, the speed of the leaves, and the planned positions. The goal was to see if the model could guess where the leaves would actually end up when the machine fired. The results were striking. The new model, which they call BKA-BPNN, predicted the leaf positions with an average error of just 0.0026 millimeters. To put this in perspective, the error was so small that the predicted position was almost indistinguishable from the actual position, with a correlation so high it was nearly perfect. When they compared this to older, standard neural networks, the new model reduced the error by nearly 90 percent. It also performed significantly better than other advanced methods that use different optimization techniques, cutting the error by almost 40 percent compared to the next best alternative.

Beyond just being more accurate, the new model was also much faster. While some competing systems took over an hour to process the data, the Black-winged Kite model completed its task in less than a minute. This speed is crucial because it means the system could potentially be used in real-time or just before a treatment to check for errors. The researchers found that the model worked consistently well across different leaf numbers and different types of treatment plans, suggesting it is a robust tool that does not need to be retrained for every single leaf or every new patient. In contrast, older models often struggled when asked to predict leaves they hadn't been specifically trained on, requiring separate training for each one. The new system handled all of them with the same high level of precision.

The study also looked at how these predictions could improve the safety of radiation therapy. Currently, doctors often verify the machine's performance after the treatment is done, which is too late to fix any mistakes. This new method offers a way to predict and correct errors before the beam ever turns on. By identifying potential deviations early, the system could help ensure that the dose delivered matches the dose planned, protecting patients from unintended exposure. The researchers noted that while their model is highly effective, it was tested on data from two specific types of radiation machines. They suggest that future work should include data from a wider variety of machines to ensure the model works everywhere. For now, however, the study demonstrates that by combining a smart search algorithm with a learning network, it is possible to predict the movements of these critical medical devices with a level of precision that was previously out of reach, offering a promising step forward in making cancer treatment safer and more effective.