Prostate cancer is typically treated with radiation directed at the entire prostate gland. However, research has shown that a specific region called the dominant intraprostatic lesion (DIL) -- usually less than 10% of the prostate's volume -- is the most common site of tumor recurrence after radiation treatment.
Studies suggest that delivering a higher radiation dose specifically to the DIL can improve tumor control without necessarily increasing harm to surrounding tissue. This approach is called dose escalation, and it requires knowing exactly where the DIL is located within the prostate at the time of each treatment session.
Multiparametric MRI (mpMRI) is the gold standard for identifying the DIL because of its superior soft tissue contrast inside the prostate. However, radiation is delivered using CT or cone-beam CT (CBCT) images, which have poor soft tissue contrast. Accurately aligning the MRI (showing the tumor) with the treatment CBCT (used during radiation delivery) is therefore a critical and technically difficult challenge.
Image registration is the computational process of aligning two images taken at different times or with different imaging technologies so that structures in one image correspond correctly to the same structures in the other. For prostate cancer, aligning MRI with CBCT is especially difficult because the two imaging modalities produce very different-looking images.
Key obstacles include: (a) differences in image appearance and intensity between MRI and CBCT; (b) prostate position and shape changes caused by varying bladder or rectal filling between imaging sessions; (c) poor soft tissue visibility on CBCT; and (d) image artifacts and noise from the CBCT acquisition process itself. Most existing approaches are limited to simple rigid registration (shifting and rotating without deformation), which cannot account for shape changes.
Deformable image registration -- which also allows stretching and warping -- is needed to handle these shape changes. One established approach uses finite element (FE) modeling to simulate how soft tissue deforms based on its physical properties. While physically realistic, FE modeling is labor-intensive and too slow for routine clinical use.
Researchers at Emory University developed a novel registration framework that combines deep learning with biomechanical constraints. The core idea is to use finite element modeling to generate physically realistic prostate deformation examples during training -- teaching the neural network how tissue should physically behave -- so that the network can then apply this knowledge instantly on new patients.
The framework uses two separate convolutional neural networks (CNNs) to automatically segment (outline) the prostate from both MRI and CBCT images. After an initial rigid alignment, the prostate shapes from both images are converted into a 3D representation called a point cloud (PC) -- essentially a set of 3D coordinate points sampling the prostate surface and interior.
A specialized Point Cloud Matching Network (PCMN) was then trained to predict how the MRI prostate point cloud should move and deform to align with the CBCT prostate point cloud. The PCMN was trained using motion vectors generated by FE simulations, so it learned to produce physically realistic deformations without needing to run an FE simulation at test time.
Point clouds are inherently different from regular image data -- they are unordered sets of 3D points without a fixed grid structure. The PCMN uses specialized point cloud convolution operations that can handle this irregular, spatially unordered data to extract meaningful features about prostate shape and position.
Each prostate was represented by 1024 sampled points, with each point encoded by its 3D coordinates relative to both the MRI and CBCT prostate centroids, plus an indicator of whether the point lies on the prostate surface. This rich encoding helps the network understand spatial relationships between the two prostates.
The PCMN was trained with two loss functions simultaneously: a supervised volumetric motion loss that penalizes predictions differing from FE-generated motion vectors, and an unsupervised surface distance loss (using Chamfer distance) that encourages the network to align prostate surfaces even for cases outside the training distribution.
The FE modeling used realistic prostate tissue properties (Young's modulus of 5 kPa and a Poisson's ratio of 0.49) modeled as a neo-Hookean hyperelastic material, which captures the nonlinear elastic behavior typical of biological soft tissues. Each FE simulation generated 10 intermediate deformation steps, multiplying the available training data from 50 patients to 500 data pairs.
The PCMN was evaluated on 50 prostate cancer patients using three metrics: Dice Similarity Coefficient (DSC) measuring prostate volume overlap, Mean Surface Distance (MSD) measuring prostate boundary agreement, and Target Registration Error (TRE) measuring internal registration accuracy using implanted fiducial markers.
The PCMN achieved a mean DSC of 0.93, MSD of 1.66 mm, and TRE of 2.68 mm. Compared to rigid registration alone, the PCMN reduced TRE by 29.1%. It also outperformed two competing deformable methods: coherent point drifting (CPD) surface registration and modality-independent neighborhood descriptor (MIND) registration.
Critically, while the CPD method achieved slightly better prostate surface alignment (MSD of 1.58 mm vs. 1.66 mm), the PCMN achieved better internal accuracy (TRE of 2.68 mm vs. 3.13 mm for CPD). This is because surface-based methods can align the outer boundary without correctly registering the interior, whereas the PCMN explicitly models internal deformation using biomechanical knowledge.
Analysis of Jacobian determinants and strain tensors confirmed that the PCMN produces physically realistic deformations that closely match the FE ground truth. This is important because unrealistic deformations (such as tissue folding or implausible stretching) could misrepresent where the DIL actually is during treatment.
Accurate MRI-CBCT registration is a foundational step for online adaptive radiotherapy, where treatment plans are adjusted before each fraction based on the patient's current anatomy. By enabling reliable DIL delineation on CBCT, this method could make dose escalation to the DIL a routine clinical practice rather than a complex research procedure.
For proton therapy specifically -- which requires more precise dose delivery than standard photon radiation -- accurate registration is especially valuable. Prior studies have shown proton therapy can achieve superior DIL boost dose distributions, but only if the tumor location is known precisely during each treatment session.
The trained PCMN completes the registration in under 1 second, making it suitable for real-time clinical workflows. This is a dramatic improvement over FE modeling, which is too slow for routine use, and represents the key practical advantage of replacing physics simulation with learned neural network inference.
The main limitation of the current method is its dependence on accurate automatic prostate segmentation. When CBCT image quality is poor -- due to photon scattering artifacts or patient movement -- the segmentation may be inaccurate, which degrades registration quality. Manual contour correction may be necessary in some cases.
Patients undergoing androgen deprivation therapy (ADT) may experience significant prostate shrinkage during their treatment course, which could challenge the registration framework. Additionally, irregular bowel or bladder filling can cause prostate deformation beyond what was represented in the training data.
Future improvements include incorporating intraprostatic fiducial marker detection directly into the registration pipeline for better initialization, expanding the training dataset to cover more diverse prostate deformation scenarios, and potentially extending the method to other pelvic cancers requiring MRI-CT registration.