Deep learning-based image registration: Robustness in adaptive radiotherapy

Medical image registration is crucial for monitoring diseases and providing image guidance during treatments, such as surgery and adaptive radiotherapy. For instance, it can transfer important structures (delineations) from the planning phase to the treatment in MRI-guided online adaptive radiotherapy (MRgRT), providing crucial information to adapt the treatment plan to the daily anatomy. In recent years, deep learning has enabled fast (sub-second) image registration. However, clinical implementation remains limited. Most deep learning-based registration methods are developed and evaluated on data from the same distribution, and their performance may not be reliable when faced with variations not represented in their training data, such as differences in image acquisition parameters or patient populations. This thesis addressed several unresolved challenges related to the robustness of deep learning registration models. The first half of this thesis focused on model robustness in broader medical image analysis contexts (Chapters 2 and 3). In Chapter 2, we introduced a concept for assessing model robustness in the absence of ground truth or expert annotations. This method builds on the concept proposed by Prasad et al. to identify a model's optimal input, i.e., it produces outputs with higher image similarity. When the optimal input deviates from the original input, it indicates susceptibility to domain shifts. This is measured using the distances, d_M and d_phi, which quantify changes in the input images and the output deformations. Our results showed that the learnable mapping improved anatomical alignment in many cases. This offers a promising initial indication that input modifications could serve as indicators of registration accuracy. The feasibility study revealed the potential to detect harmful simulated domain shifts, as higher values of d_phi were associated with poorer registration accuracy. However, its ability to identify subtle performance degradations in real-world scenarios was not demonstrated. Additionally, the hypothesis that the learnable image mapping would produce perceptible and informative changes was not supported; the modifications were subtle and visually imperceptible. These findings suggest that further refinement is needed to achieve more interpretable image mappings. In Chapter 3, we investigated transfer learning as a strategy to enhance model robustness, pioneering this approach in the context of registration models. We first pre-trained a generic registration model on synthetic data and then fine-tuned it for specific target domains. To evaluate robustness, we tested the fine-tuned model on previously unseen brain MRI and lung CT datasets. We compared its performance against models trained from scratch, both with and without data augmentation. Additionally, we studied how design choices in synthetic data affected cross-domain performance. Our results showed that transfer learning improved robustness to several unseen datasets. Specifically, it yielded accuracy improvements of 34–62% for lung CT registration and 6–14% for brain MRI registration. On other datasets, transfer learning offered only marginal gains over training from scratch. More advanced transfer learning techniques may further improve performance. In addition, while it was important to include random geometric shapes and dynamic contrast in the synthetic dataset to achieve the most consistent registration of the pre-trained model across tasks, different target tasks may benefit from different synthetic data. Our approach was data-efficient, while it did not depend on acquired medical images during pre-training and it required less domain-specific data than training from scratch. The second half of this thesis applied deep learning-based registration for MRgRT, diving into applied aspects of online re-contouring methods (Chapters 4 and 5). While the chapters involve different challenges and methodologies, they are connected through their shared emphasis on robustness to variability. In Chapter 4, we developed a registration framework for daily target and organ-at-risk (OAR) contouring in MRgRT for prostate cancer. The main contributions of this chapter are integrating both rigid and deformable registrations into a unified deep learning framework and comparing three promising model architectures for deformable registration within the framework. Specifically, we evaluated their accuracy, speed, and robustness to simulated domain shifts. Our results showed that the Laplacian image registration network (LapIRN) was the most accurate and robust among the studied deformable networks. LapIRN demonstrated superior robustness against large nonlinear deformations, maintaining better contour accuracies than the other models (U-Net and MS-D Net). LapIRN's superior performance is likely due to its three-stage coarse-to-fine registration process, which aligns with other studies that have found benefits of multi-resolution, cascaded models over single networks. The resulting target (prostate) contour accuracy was close to inter-observer agreement, with a Dice score of 0.89 compared to reported inter-observer Dice scores ranging from 0.87 to 0.94. The OAR (bladder) contours were less accurate (Dice 0.86 vs. inter-observer Dice 0.93) due to challenges with bladder filling and emptying, leaving significant room for improvement. In Chapter 5, we addressed the complex task of delineating the mesorectal clinical target volume (CTV) in online adaptive MRgRT for rectal cancer. Unlike previous chapters that focused solely on registration, Chapter 5 also compared segmentation approaches. This broader perspective was chosen because segmentation models benefit from accessible, well-established frameworks such as nnU-Net. To improve contour accuracy, we incorporated the planning CTV into the models, using its shape and boundary information as anatomical guidance. This strategy differs from Chapter 4, where registration was performed using image data alone, without prior anatomical input. The inclusion of the planning CTV in Chapter 5 was motivated by the mesorectal region’s challenging anatomy, particularly at the craniocaudal boundaries, where variability often leads to segmentation inaccuracies. We evaluated the methods' accuracy, clinical usability, and robustness. Incorporating the planning CTV improved the contour accuracy of both models (Segmentation Hausdorff distance (HD): 17.5 → 9.3 mm; Registration HD_middle: 7.1 → 6.0 mm). Comparing the segmentation and registration models, segmentation achieved slightly higher accuracies in the mid and cranial regions of the CTV and produced more clinically acceptable contours (9/10 versus 3/10 for registration). The segmentation model was also slightly more robust to rectal filling differences between planning and daily treatment. However, its robustness was not flawless as accuracy decreased with 6 mm translations and with 30% rectal emptying, underscoring the importance of robustness assessments.