Making deep learning reliable for adaptive radiotherapy

2 december 2025

Medical images are essential for accurate diagnosis and treatment, such as surgery and radiotherapy. To make these images useful, they often need to be precisely aligned. Deep learning, a smart form of artificial intelligence, makes this process much faster. But how reliable is it when the data changes? PhD candidate Iris Kolenbrander explored this question and found that speed does not always mean robustness. She successfully defended her PhD at the Department of Biomedical Engineering on December 2.

Photo by Vincent van den Hoogen

Doctors use medical images to detect diseases and plan treatments. Often, different images of a patient must fit together perfectly. This process, called image registration, allows doctors to track changes over time, combine images from different techniques, and locate tumors accurately for treatment.

Traditional methods are precise but slow, taking minutes to align images. Deep learning can do this in seconds, saving valuable time in clinical settings. However, deep learning models learn from large datasets. If new images differ from those used during training, for example due to different scanners or patient populations, accuracy can drop. This poses a risk for patient safety.

Cover of the thesis bij Iris Kolenbrander

How robust is deep learning when data changes?

Kolenbranders鈥� research focused on testing how well deep learning models perform when medical data varies. This is important because changes in patient anatomy or differences in imaging equipment can affect accuracy. She examined two main areas:

General medical image analysis

MRI-guided radiotherapy (MRgRT), where treatment plans are adapted daily to match the patient鈥檚 anatomy

MRgRT allows radiation to be delivered with high precision, but it requires daily image registration. That means both speed and reliability are critical for safe treatment.

Key findings

Detecting problems: Kolenbrander developed a method to check if a model still performs well without needing a perfect reference. This method could detect major issues but struggled with subtle errors.
Transfer learning helps: Pre-training models on synthetic data and then fine-tuning them improved robustness for new datasets.
Prostate cancer: A deep learning framework for automated tumor contouring achieved accuracy close to human experts.
Rectal cancer: Segmentation, which directly outlines the tumor, proved more accurate and clinically acceptable than registration, although both were sensitive to anatomical changes.

Impact on healthcare

Deep learning makes medical image registration faster and more efficient, but clinicians must remain cautious. Techniques like transfer learning can help improve reliability. Kolenbranders鈥� research shows that AI has great potential in healthcare, but continued collaboration between researchers and medical professionals is essential to ensure safety and robustness.