Transrectal ultrasound (TRUS) is one of the most widely used imaging methods for guiding prostate cancer diagnosis and treatment. It is economical, portable, and provides real-time imaging, making it indispensable for procedures like biopsy guidance and radiation treatment planning. However, automated analysis of TRUS images is persistently difficult.
TRUS images are notoriously challenging because they contain speckle noise, shadow artifacts, and low contrast between the prostate and surrounding tissue. The prostate boundary can appear blurred, intensity distributions are uneven, and the shape of the gland varies substantially between patients. These factors cause significant inaccuracy in automated segmentation.
Accurate prostate segmentation is clinically important for multiple reasons: it allows precise volume measurement, enables fusion with MRI for targeted biopsy, and is fundamental to radiotherapy planning. Currently, manual segmentation by specialists is time-consuming, tedious, and subject to inter-observer variability.
A major bottleneck for training deep learning segmentation models is the need for large amounts of annotated patient data. Prostate cancer (PCa) is less common than benign prostatic hyperplasia (BPH), meaning far more ultrasound images of BPH patients exist in clinical databases. This study investigated whether a segmentation model trained on BPH images could be directly applied to PCa images, potentially resolving the data shortage problem.
The study collected 370 ultrasound images from 260 BPH patients and 68 ultrasound images from 62 PCa patients, all confirmed by prostate biopsy. The BPH images were divided into a training set (approximately 80%) and a test set (approximately 20%), while all PCa images were used exclusively as a test set, never seen during training.
Before building segmentation models, the researchers first confirmed that PCa and BPH prostate images are measurably different. Using radiomics, a technique that extracts hundreds of quantitative features from medical images, 924 features per image were calculated. LASSO regression was used to identify the most discriminative features, selecting 9 features that distinguished PCa from BPH, including shape elongation, sphericity, texture intensity patterns, and wavelet-based features.
Three deep learning architectures were then trained exclusively on BPH data: U-Net, LinkNet, and PSPNet. U-Net uses a symmetric encoder-decoder structure with skip connections that preserve spatial detail across different levels of abstraction. LinkNet also uses an encoder-decoder approach, while PSPNet uses pyramid pooling to capture features at multiple scales.
All models were trained with 2,000 augmented images using a learning rate of 0.001, batch size of 32, and 100 epochs. After segmentation, four post-processing strategies were tested: no post-processing, gap-filling, convex hull closure, and selection of the largest connected region, to assess which best handled the prostate's irregular boundaries.
Among the three architectures, U-Net achieved the highest Intersection over Union (IoU) score of 0.9602 on the BPH validation set, significantly outperforming LinkNet and PSPNet. IoU, also called the Jaccard index, measures the overlap between the predicted segmentation and the ground truth annotation, with 1.0 being a perfect match.
The critical finding was that the U-Net model trained on BPH images performed comparably when applied to PCa images. Independent-sample t-tests showed no statistically significant difference (p > 0.05) between the Dice coefficients achieved by the model on BPH test images versus PCa test images. This held true across all four post-processing methods tested.
The model maintained performance even when images were resized to match the original ultrasound aspect ratio rather than the standard 512x512 pixels, confirming that image depth variation from different ultrasound probe positions had minimal impact on segmentation accuracy. This robustness to image dimension changes is important for clinical deployment where image parameters vary.
The researchers validated this result using two-tailed statistical tests after confirming data normality with the Shapiro-Wilk test. The statistical equivalence confirms that, from a segmentation perspective, the prostate in BPH and PCa ultrasound images is similar enough that training data from BPH patients can be used to build models applicable to cancer patients.
Despite the overall segmentation success, the radiomics analysis confirmed that BPH and PCa images are not identical. The 9 discriminative features selected by LASSO regression reveal meaningful differences in how the prostate appears in the two conditions.
Shape features were among the key discriminators. The elongation feature (ratio of the two largest shape dimensions) and sphericity feature (ratio of tumor perimeter to the perimeter of an equal-area circle) differed between conditions, suggesting that the prostate in PCa tends to have a more pronounced spherical shape component than in BPH.
Texture features derived from gray-level run-length matrices and wavelet transforms also distinguished the two conditions, indicating differences in tissue heterogeneity and multi-scale intensity patterns between benign and malignant prostatic tissue. These radiomic differences are consistent with known histological differences between BPH (enlarged glandular and stromal tissue) and PCa (invasive tumor cells disrupting normal architecture).
The fact that these measurable differences exist yet the segmentation model still transfers successfully suggests that what matters for accurate boundary detection is not the internal texture but the overall prostate boundary shape and location, which appears similar enough between BPH and PCa to allow cross-condition model transfer.
The central clinical importance of this study is its potential solution to a fundamental data scarcity problem in medical AI development. Training high-precision segmentation models typically requires thousands of annotated images from the disease of interest, but annotated PCa ultrasound data is inherently scarce compared to BPH data.
By demonstrating that BPH-trained models can segment PCa prostates with equivalent accuracy, this work opens the possibility of using the much larger repositories of annotated BPH ultrasound images available at urology departments worldwide to bootstrap PCa segmentation systems without requiring re-annotation of cancer-specific datasets.
This has practical implications for lower-resource healthcare settings where PCa case volumes are insufficient to train dedicated AI models but BPH data is plentiful. A model trained on BPH data at a large center could potentially be deployed for PCa segmentation at smaller clinics, extending access to AI-assisted prostate imaging tools more broadly.
The study also highlights the importance of radiomics as a complementary tool to deep learning. Even when deep learning segmentation is the primary output, radiomic feature analysis can identify imaging biomarkers that distinguish cancer from benign disease, which could support diagnosis or risk stratification alongside segmentation.
This study demonstrated that a U-Net model trained entirely on BPH ultrasound images can be directly applied to PCa prostate segmentation with statistically equivalent performance. The peak IoU of 0.9602 on BPH data and the preserved accuracy on PCa data represent clinically useful segmentation quality.
The study is limited by its relatively small PCa test set (68 images from 62 patients), which provides statistical equivalence but would benefit from larger-scale validation. Additionally, the single-center design means performance in other clinical environments, with different ultrasound equipment and operator protocols, needs to be confirmed.
Future work should explore whether BPH-trained models can also support downstream tasks beyond segmentation, such as volume estimation, lesion detection, or biopsy targeting, which all depend on accurate prostate boundary delineation.
The broader principle established here, that models trained on abundant related-disease data can transfer to scarce target-disease data when the structural features are sufficiently similar, may apply beyond prostate imaging to other cancer sites where training data is limited.