Traditional radiotherapy planning for prostate cancer uses computed tomography (CT) scans to define radiation dose because CT provides the electron density information needed for dose calculation. However, CT offers inferior soft-tissue contrast compared to magnetic resonance imaging (MRI), making it harder to precisely delineate the prostate and surrounding structures on CT alone.
MRI-only planning eliminates the need for CT by using MRI for both anatomy delineation and, through synthetic CT generation, dose calculation. This approach reduces radiation exposure from CT, simplifies the imaging workflow, avoids registration errors that arise when MRI and CT images are fused, and is more patient-friendly.
A critical step in MRI-only planning is auto-segmentation: automatically delineating the target structures (the prostate and clinical target volume) and organs-at-risk (bladder, rectum, femoral heads, urethra) directly on MRI. Manual delineation is time-consuming and varies between observers, making accurate and fast automated segmentation essential for clinical adoption of MRI-only workflows.
The study compared two deep learning architectures for automated segmentation of six structures on T2-weighted MRI: the clinical target volume (CTV), bladder, rectum, left femoral head, right femoral head, and urethra. The CTV in prostate cancer typically encompasses the prostate gland and potentially the proximal seminal vesicles.
DeepLabV3+ is a semantic segmentation architecture originally developed for natural image understanding. It uses atrous (dilated) convolutions to capture multi-scale context without reducing spatial resolution, and incorporates an encoder-decoder structure with skip connections. The model was initialized using weights pre-trained on natural images from the ImageNet dataset, a technique called transfer learning.
U-Net is the most widely used deep learning architecture in medical image segmentation. It features a symmetric encoder-decoder structure with skip connections that preserve spatial detail, and was trained from scratch on the prostate MRI dataset. The comparison between a transfer-learning approach and a from-scratch medical imaging architecture is directly relevant to practical deployment decisions.
The dataset comprised 100 patients split into training and test sets. All patients had T2-weighted MRI scans and manual contours drawn by experienced radiation oncologists, which served as the reference standard for evaluating automated segmentation accuracy.
DeepLabV3+ significantly outperformed U-Net for five of the six segmented structures (P less than 0.001 for all), with the exception of the urethra where performance was comparable. The primary metric used was the volumetric Dice Similarity Coefficient (VDSC), a standard measure of overlap between the automatically generated contour and the manual reference contour.
DeepLabV3+ achieved a VDSC of 0.93 for the bladder and 0.83 for the CTV, which are considered clinically acceptable levels of overlap. The bladder result is particularly strong given that bladder filling can vary considerably between patients and sessions.
The urethra was the most challenging structure for both methods, reflecting its small size, lack of intrinsic contrast on T2-weighted MRI, and the fact that it must be inferred from anatomical context rather than directly visualized as a distinct structure. Both networks struggled comparably with this structure, suggesting that the urethra may require dedicated approaches or multi-sequence input.
Transfer learning leverages a neural network's weights that were learned on a large dataset for a different task. In this case, DeepLabV3+ was first trained on the ImageNet dataset containing millions of natural photographs across 1,000 categories. This pre-training taught the network to detect low-level features like edges and textures as well as higher-level structural patterns.
When fine-tuned on the much smaller prostate MRI dataset, these pre-learned feature detectors provided a strong starting point that required fewer medical images to reach high performance. U-Net, trained from scratch on only the 100-patient medical dataset, had to learn all features entirely from the limited available data.
Medical image datasets are inherently small compared to natural image datasets because acquiring images requires clinical infrastructure, patient consent, and expert annotation. Transfer learning from natural image datasets addresses this fundamental data scarcity problem in medical AI, and the superior performance of DeepLabV3+ in this study provides direct evidence for its practical value in prostate MRI segmentation.
The results support integrating DeepLabV3+ as an auto-segmentation tool in clinical MRI-only prostate radiotherapy planning workflows. A VDSC of 0.83 for the CTV means that approximately 83% of the volume delineated by the automatic system overlaps with what the radiation oncologist would draw manually, with the automated contour serving as a starting point that requires only minor editing.
In routine clinical practice, auto-segmentation tools are not expected to be perfect; they are expected to reduce the time radiation oncologists spend on contouring and to provide consistent, reproducible starting points. Even a tool requiring modest manual editing can reduce contouring time from 30-60 minutes per patient to 5-10 minutes, a meaningful efficiency gain in high-volume centers.
The comparable performance of both methods for the urethra highlights a remaining clinical gap: the urethra is an important organ-at-risk in prostate radiotherapy because urethral dose is associated with urinary side effects. Developing reliable automated urethra segmentation remains an important unsolved problem.
This study provides evidence that transfer learning-based deep learning, specifically DeepLabV3+ pre-trained on ImageNet, can achieve clinically acceptable automated segmentation of prostate cancer treatment structures on T2-weighted MRI, outperforming the widely used U-Net architecture trained from scratch.
The findings support a pathway toward fully automated or minimally supervised MRI-only prostate radiotherapy planning. Combined with advances in synthetic CT generation for dose calculation, such approaches have the potential to streamline the treatment planning workflow, reduce inter-observer variability, and improve patient access to high-quality radiotherapy.
Future research should address the remaining challenge of urethra segmentation, investigate performance in patients with implants or unusual anatomy, and evaluate whether adding DWI or other MRI sequences as additional input channels could further improve segmentation accuracy for difficult structures.