Prostate cancer is one of the most common cancers among men, and Intensity-Modulated Proton Therapy (IMPT) offers a powerful treatment option because it can deliver a highly localized radiation dose directly to the tumor while minimizing damage to surrounding healthy tissues such as the bladder and rectum.
Unlike conventional photon-based radiation, proton therapy is highly sensitive to daily changes in a patient's anatomy. The prostate, bladder, and rectum can shift position or change shape between treatment sessions due to factors like bowel gas, bladder filling, or rectal distension, which can significantly alter how the radiation dose is actually delivered.
To account for these variations, online adaptive radiotherapy adjusts the treatment plan at each session based on imaging taken on that day. This requires fast and accurate automatic delineation of target structures and organs-at-risk, a process that is too slow if done manually every day.
The researchers developed a multi-step pipeline combining deep learning and deformable image registration to automatically transfer contours drawn on a planning CT scan to a new daily CT scan. The pipeline was built around elastix software, a widely used open-source registration framework.
The first stage involved training a 3D Convolutional Neural Network (CNN) to automatically segment the bladder in each CT scan. Bladder shape and volume vary considerably day to day, and using an accurate bladder segmentation as an anatomical constraint greatly improves registration accuracy compared to relying on image intensity alone.
A Generative Adversarial Network (GAN) was also incorporated to handle cases where large gas pockets appear in the rectum, which can make image registration unreliable. The GAN was trained to synthetically fill in or inpaint these gas regions so that registration algorithms are not misled by the unusual image contrast they create.
Three datasets from different institutions and CT scanner manufacturers were used: one for training and testing the CNN, and two additional datasets from separate institutes for independent validation of the full pipeline. This multi-site design was critical to demonstrating that the system would generalize beyond the data it was trained on.
Deformable image registration (DIR) works by computing a non-rigid transformation that aligns one image to another, allowing contours from the original planning image to be mapped onto the new daily image. However, standard intensity-based registration can fail when large, unpredictable anatomical changes occur, particularly in the rectum.
By incorporating the automatic bladder segmentation as an additional input to the registration, the algorithm gains explicit anatomical knowledge. The optimizer simultaneously minimizes differences in image intensity and differences in bladder shape, which anchors the registration around a well-defined structure and reduces errors in adjacent organs.
The gas inpainting step using the GAN addresses a well-known failure mode: when a large gas pocket suddenly appears or disappears in the rectum, the registration algorithm may incorrectly deform the surrounding anatomy to match. Replacing gas pockets with synthetic tissue-like content removes this source of confusion, improving robustness significantly.
The bladder segmentation CNN achieved a Dice Similarity Coefficient (DSC) of 88% on the primary test dataset and 82% on the independent validation dataset, demonstrating reliable performance across different CT scanner types and imaging centers.
For the registration pipeline, the Mean Surface Distance (MSD) between automatically propagated contours and manually drawn contours was 1.29 mm for the prostate, 1.48 mm for the seminal vesicles, and 1.49 mm for lymph nodes on the second dataset. On the third dataset, MSD values were 2.31 mm for the prostate and 1.76 mm for the seminal vesicles, reflecting greater anatomical variability at that institution.
These geometric accuracy figures are clinically meaningful: sub-2 mm errors for most structures indicate that the automatically generated contours closely match what a radiation oncologist would draw manually, which is a prerequisite for using them in treatment planning without routine correction.
Geometric accuracy alone does not guarantee that automatically generated contours are safe to use in radiation treatment planning. The researchers therefore validated the contours by actually generating IMPT treatment plans based on the propagated outlines and checking whether those plans met all required dosimetric constraints.
A contour was considered clinically acceptable if the treatment plan generated from it satisfied dose coverage requirements when evaluated against the manually drawn reference contours. This is a rigorous real-world test because a contour that looks geometrically close may still produce an unacceptable plan if the deviation occurs at a dosimetrically sensitive location.
The automatically propagated contours produced clinically acceptable plans in 86% of prostate cases, 91% of seminal vesicle cases, and 99% of lymph node cases. These figures directly translate into how often a radiation therapist would need to manually correct the automated contours before proceeding with treatment.
The study introduced the Conservative Success Rate (CSR) as its primary composite metric. The CSR is defined as the fraction of cases where all structures simultaneously meet the clinical acceptability criteria, making it a strict measure of overall pipeline robustness.
The proposed combined pipeline achieved a CSR of 80%, meaning that in 4 out of every 5 treatment fractions, the fully automatic system produced plans that required no manual correction at all. This compares favorably to a CSR of 65% achieved by intensity-based registration alone, representing a substantial improvement in reliability.
The 15-percentage-point gain in CSR is clinically significant because it means the burden on radiation therapists during an adaptive treatment session is substantially reduced. In high-volume centers treating dozens of patients per day, even modest reductions in manual correction time per patient can make online adaptation feasible in routine practice.
The study demonstrates that combining convolutional neural network-based segmentation, GAN-based image synthesis, and anatomy-constrained deformable registration can achieve the accuracy and robustness needed for practical online adaptive proton therapy in prostate cancer.
The ability to automatically propagate contours with 80% success rate without any manual intervention represents a meaningful step toward making online adaptive IMPT clinically feasible. Prior limitations in registration robustness had been a major barrier to widespread adoption of this approach.
Future work should focus on further reducing the remaining 20% failure rate, particularly for patients with extreme rectal gas or unusual anatomy, and on validating the system in prospective clinical trials. The methodology is also potentially extensible to other anatomical sites where daily anatomical changes challenge standard radiotherapy delivery.