Radiotherapy for prostate cancer requires extreme precision: delivering a lethal dose to the tumor while protecting nearby structures like the rectum, bladder, and femoral heads. Planning CT (pCT) scans are the gold standard for this, providing accurate tissue density information measured in Hounsfield Units (HU) -- a scale where water is 0 HU, air is -1000 HU, and dense bone is around +1000 HU. These values are essential for calculating how radiation will interact with and pass through each tissue type.
During a typical radiotherapy course lasting several weeks, the patient's anatomy changes daily -- the bladder fills or empties, the rectum expands with gas, and the prostate itself can shift position. Cone Beam CT (CBCT) is acquired directly on the treatment table before each daily fraction, allowing clinicians to verify the patient's position and detect anatomical changes in real time. However, CBCT images suffer from significant scatter artifacts, poor soft-tissue contrast, and -- critically -- inaccurate HU values due to differences in imaging geometry and the lower radiation dose used for acquisition.
Because CBCT HU values are unreliable, it cannot be used directly for dose recalculation. This means that even when clinicians detect a significant anatomical change on CBCT -- for example, a grossly distended rectum pushing the prostate out of the planned treatment volume -- they often cannot immediately recalculate the dose and adapt the plan without acquiring a new planning CT.
Adaptive radiotherapy (ART) aims to address this by adjusting treatment plans based on actual daily anatomy rather than a static plan made weeks earlier. But online ART requires fast, accurate dose calculation from images available on the treatment table -- a need that CBCT cannot currently fulfill due to its HU inaccuracy.
Synthetic CT (sCT) generation converts a CBCT image into a CT-like image with corrected, physically accurate HU values. If the sCT is sufficiently accurate, it can be used directly for dose recalculation and treatment plan adaptation -- enabling online ART without requiring an additional planning CT scan.
Several approaches to sCT generation have been proposed. Bulk density assignment manually maps regions to fixed HU values. Atlas-based methods warp a database of CT images to match the patient's CBCT anatomy. Generative adversarial networks (GANs) can produce high-quality images but require large datasets, long training times, and substantial computational resources, limiting their feasibility in routine clinical settings.
Deep learning-based CNN methods, particularly U-Net architectures, have emerged as practical alternatives. They can be trained on paired CBCT-CT datasets and learn to map CBCT appearance directly to corrected HU values. Prior studies on pelvic CBCT have shown reductions in mean absolute HU error from 92 HU to as low as 31 HU using U-Net, and some GAN-based studies have achieved errors around 8 HU -- demonstrating the range of what is technically achievable.
This study focuses on two lightweight architectures -- U-Net and ResU-Net -- as practical tools for prostate cancer ART. The emphasis on lightweight design addresses a real clinical gap: many radiotherapy departments lack the computational infrastructure to run large GAN models in real time, and a simple, effective architecture that works well on limited hardware is often more clinically useful than a theoretically superior model that cannot be deployed.
This retrospective study included 10 prostate cancer patients treated at a single institution between 2019 and 2024. For each patient, one planning CT and five CBCT scans acquired on different treatment days were collected, yielding 50 total CBCT-CT paired image sets. The multi-temporal design is clinically important: it reflects the real scenario where the same patient's anatomy varies across fractions, testing model robustness to these daily changes.
Images were preprocessed through rigid registration to align CBCT volumes with corresponding planning CT slices, followed by visual verification by a radiation oncologist and medical physicist. Slices with severe mismatches due to bladder or rectal filling differences were excluded. The final dataset retained 1,890 slice pairs, split into 1,323 training, 283 validation, and 284 test slices. All images were standardized using z-score normalization before model input.
U-Net uses a classic encoder-decoder design with skip connections that pass feature maps from encoder to decoder at each resolution level, preserving both local detail and global context. The encoder progressively compresses the image through four blocks (32 to 512 channels), and the decoder reconstructs the sCT output at full resolution. A dropout rate of 0.5 in the bottleneck layer reduces overfitting.
ResU-Net adds residual connections (identity shortcuts) within each convolutional block. These shortcuts allow gradients to flow more easily during training and stabilize learning in deeper networks, enabling more effective feature extraction without adding significant model complexity. Both models were trained using a combined L1 and MSE loss function to balance structural similarity with HU accuracy.
All three quality metrics improved significantly from raw CBCT to both sCT models. Structural Similarity Index (SSIM) measures how closely two images match in structure, luminance, and contrast (1.0 = identical). CBCT achieved a mean SSIM of 0.763 against the planning CT, which improved to 0.840 with U-Net and 0.851 with ResU-Net -- both statistically significant improvements. The difference between U-Net and ResU-Net was not statistically significant for SSIM.
Peak Signal-to-Noise Ratio (PSNR) measures image signal fidelity in decibels (higher is better). CBCT achieved 21.55 dB, which improved to 24.74 dB with U-Net and 25.24 dB with ResU-Net. ResU-Net's advantage over U-Net was statistically significant (p less than 0.05). The ResU-Net value of 25.24 dB falls within the 25-28 dB range cited in the literature as acceptable for clinical sCT quality.
Mean Absolute Error (MAE) in HU directly quantifies how far synthetic CT tissue density values deviate from the reference planning CT. CBCT had an MAE of 75.2 HU. U-Net reduced this to 65.3 HU and ResU-Net to 61.8 HU -- both statistically significant improvements, with ResU-Net also significantly outperforming U-Net. These values are higher than GAN-based results (approximately 8-58 HU in prior studies) but represent a meaningful improvement for a lightweight model trained on only 10 patients.
Qualitative difference maps showed that CBCT artifacts were most pronounced at bone-to-soft-tissue interfaces, particularly around the femoral heads. Both sCT models substantially reduced these deviations, with ResU-Net producing tighter HU distributions and fewer outliers. The visual quality of ResU-Net synthetic CTs was notably closer to planning CT in appearance, particularly in the critical pelvic anatomy region encompassing the prostate.
The primary clinical motivation for this work is enabling online adaptive radiotherapy (ART) -- adjusting the radiation plan during a treatment session based on real-time imaging rather than the original plan made weeks earlier. For prostate cancer specifically, daily variations in rectal and bladder filling can shift the prostate by several millimeters, potentially moving the tumor out of the high-dose region while exposing healthy tissue to more radiation than planned.
If sCT images with sufficient HU accuracy can be generated from daily CBCT in seconds or minutes, radiation physicists could use them to recalculate the dose distribution and confirm whether the current plan remains acceptable or requires adaptation -- all before the patient leaves the treatment table. This capability is the core promise of ART-compatible sCT generation.
The MAE values achieved here (61-65 HU) are higher than the sub-20 HU accuracy typically targeted for proton therapy, where HU errors of even a few units can translate to millimeter range shifts. For photon-based prostate radiotherapy, where dosimetric sensitivity to HU errors is lower, these values may be clinically acceptable for adaptive plan assessment -- but formal dosimetric validation using dose-volume histograms (DVH) was not performed in this study.
A key practical advantage of the lightweight architecture approach is deployability on standard clinical hardware. This study was conducted on a consumer GPU (NVIDIA GeForce GTX 1660, 6GB memory), which is far less powerful than the high-end workstations typically required for GAN training and inference. Demonstrating feasibility on accessible hardware makes the approach immediately relevant to community hospitals and institutions without specialized AI infrastructure.
The most significant limitation is the small dataset: 10 patients and 50 CBCT-CT pairs. While sufficient for a feasibility study, this is far too small to establish robust clinical generalizability. Performance will need to be validated on larger, multi-institutional datasets with diverse scanner hardware, patient body habitus, and clinical protocols before clinical deployment.
The models operate on 2D slices independently, treating each image as a separate input without three-dimensional context from adjacent slices. This design is simpler to train but misses spatial relationships between consecutive slices that can improve consistency and reduce artifacts in sCT volumes. Future work should explore 3D CNN architectures that process full volumetric inputs.
The study evaluated image quality only -- SSIM, PSNR, and MAE. Clinical deployment requires demonstrating that sCT-based dose calculations produce dose-volume histograms (DVH) that match reference CT-based calculations within clinically acceptable tolerances, and that auto-contouring algorithms can accurately segment pelvic structures from sCT images. Neither validation was performed here.
Despite these limitations, the study provides a strong proof-of-concept that simple, accessible deep learning architectures can meaningfully improve CBCT image quality for prostate radiotherapy. The multi-temporal design -- using CBCTs from different treatment fractions of the same patients -- is a more realistic test of clinical applicability than studies that use a single fraction per patient.
This study demonstrates that U-Net and ResU-Net can generate synthetic CT images from prostate CBCT with significantly improved HU accuracy and structural similarity compared to raw CBCT -- using only 10 patients and standard consumer GPU hardware. ResU-Net consistently outperformed U-Net, particularly on PSNR (25.24 dB) and MAE (61.8 HU), with its residual connections providing more stable training and better feature extraction.
The practical contribution of this work is not that it achieves the best numbers in the field -- GAN-based models and large-scale studies report better accuracy -- but that it shows these gains are achievable with simple, resource-efficient architectures in a limited-data clinical environment. For the many radiotherapy centers that lack large patient databases or high-end computing infrastructure, this is a meaningful and actionable finding.
The path to clinical deployment requires dose calculation validation, auto-contouring evaluation, and prospective studies in larger patient cohorts. These steps will determine whether the image quality improvements translate into genuine improvements in treatment delivery accuracy and patient outcomes.
As ART becomes increasingly standard in prostate radiotherapy, sCT generation from daily CBCT will be a core enabling technology. Lightweight, clinically practical deep learning models -- rather than only the most sophisticated architectures -- will be essential for making ART accessible across the full range of treatment settings, not just highly specialized academic centers.