A deep learning derived prostate zonal volume-based biomarker from T2-weighted MRI to distinguish between prostate cancer and benign prostatic hyperplasia

Abstract

Background: Benign prostatic hyperplasia (BPH) and prostate cancer (PCa) share overlapping characteristics on magnetic resonance imaging (MRI), confounding the diagnosis and detection of PCa. There is thus a clinical need to accurately differentiate BPH-Only from BPH-PCa to prevent overdiagnosis and unnecessary biopsies. Although BPH and PCa may share overlapping features, they are distinct clinical entities. Previous evidence suggests that prostate peripheral zone (PZ) and transition zone (TZ) volumes on MRI are differentially associated in patients with BPH-PCa versus those with BPH-Only.

Purpose: To develop and validate the ratio of machine learning derived PZ and TZ volumes on T2-weighted (T2W) MRI as an imaging biomarker to distinguish BPH-PCa and BPH-Only.

Methods: In this single-center, retrospective study, we identified N = 199 patients (106 BPH-Only and 93 with both BPH-PCa) who underwent a three Tesla multi-parametric MRI before systematic biopsy. A radiologist and a urologist jointly annotated PZ and TZ regions of interest on T2W, involving all 199 cases. We presented and trained a 3D conditional generative adversarial network (cGAN)-based prostate zone volume segmentation model (ProZonaNet) to segment 3D prostate TZ, PZ volumes on T2W MRI. We used 139 cases (with 7× data augmentation, yielding 973 training volumes) and an independent test set of 60 cases to train and evaluate ProZonaNet. ProZonaNet was optimized in terms of dice similarity coefficient (DSC). We then computed prostate zonal volume ratio (pZVR = TZ/PZ) from both ProZonaNet segmentations and ground-truth annotations on all 199 cases, evaluating agreement using Concordance Correlation Coefficient (CCC). The pZVR biomarker was assessed for its ability to distinguish BPH-PCa from BPH-Only. Univariate and multivariate analyses were performed to evaluate the independent effect of pZVR over clinical parameters.

Results: ProZonaNet achieved a mean mDICE of 92.5% on the independent test set (N = 60), outperforming state-of-the-art 3D segmentation models. The computed pZVR showed high agreement with ground-truth annotations, with CCC values of 0.960 for BPH-Only and 0.930 for BPH-PCa cases. The pZVR computed using ProZonaNet, along with two other clinical parameters, including age and prostate-specific antigen, improved the AUC from 0.758 to 0.927 in distinguishing between BPH-Only and BPH-PCa. At the same time, on a subset of low-grade prostate cancer cases (106 BPH-Only and 23 BPH-PCa with Gleason Score = 3+3), the integrated pZVR model improved the AUC from 0.750 to 0.910 in distinguishing between patients with BPH-Only versus BPH-PCa. On both univariate and multivariate analyses, pZVR demonstrated significant discrimination between patients with BPH-Only versus BPH-PCa.

Conclusions: We demonstrated that the prostate zonal volume ratio computed with our ProZonaNet can be used to differentiate benign prostatic hyperplasia from prostate cancer on MRI. These results demonstrate the feasibility of non-invasive diagnosis of BPH-PCa, potentially aiding in the ability to distinguish PCa from benign cancer confounders such as BPH-Only.

Keywords: MRI multi‐region 3D segmentation; cGAN; prostate cancer.

FIGURE 1

(a) Extraction of TZ&PZ in prostate MRIs using the pretrained ProZonaNet model, the detailed model architecture is illustrated in Figure 2. The input MRIs have been resampled to the fixed size of 128 × 128 × 32 voxels; (b) volume feature calculation based on the segmented results; (c) feature grouping and testing; (d) clinical validation.

FIGURE 2

The detailed schematic diagram of ProZonaNet (The input, output, and annotation volumes are represented by red, green, and blue edged cubes, respectively. The 3D discriminative maps of the discriminator for synthetic or real regions of interest are also represented by green‐ and blue‐edged cubes, respectively).

FIGURE 3

Patient selection chart and data description.

FIGURE 4

Concordance correlation coefficient (CCC) analysis for pZVR agreement between ProZonaNet‐derived values and ground truth annotations for PCa and BPH cases. The scatter plots illustrate the correlation between automatically computed and manually annotated pZVR values for BPH‐PCa cases (left) and BPH‐Only cases (right).

FIGURE 5

Prostate zonal segmentation results with ProZonaNet and comparison against other SOTA methods (From left to right, each column presents the mask from 3DU‐net, 3DResU‐net, HighRes‐net, 3DDense‐net, 3DDenseVoxel‐net, Vnet, ProZonaNet, and Ground Truth separately. Red and green represent the segmented transition and peripheral zones, respectively. The first four rows from top to bottom represent axial, Sagittal, coronal, and 3D views. The last row outlines the segmentation results of different models).

FIGURE 6

T‐test results of pZVR (pZVR was included in the multivariable logistic regression model and its added value on prostate cancer diagnosis was also estimated by the incremental value of AUC from two logistic regression models [one with and another without pZVR] through receiver operating characteristics [ROC] analysis. The optimal cutoff of the risk score to estimate the sensitivity and specificity was estimated by maximizing the Youden index. All tests are two‐sided and p values ≤ 0.05 were considered statistically significant.).

FIGURE 7

Feature importance visualization using SHAP values. The beeswarm plot illustrates the distribution of Shapley values, comparing the overall prediction importance of TZ_PZ and clinical variables across all 199 patients (106 BPH‐Only and 93 BPH‐PCa). Each point represents a variable's contribution to the prediction for an individual patient or sample. The plot demonstrates how each variable influences the model's output, either increasing or decreasing the prediction compared to the average.

FIGURE 8

Inclusion of pZVR with clinical parameters (age, tPSA) improves distinguishing BPH‐PCa and BPH‐Only in comparison to clinical parameters alone. ROC curves from two models: TZ_PZ (pZVR) + age + tPSA (AUC = 0.927) and age + tPSA (AUC = 0.758).

FIGURE 9

Feature importance analysis through SHAP values for distinguishing clinically insignificant BPH‐PCa cases. The beeswarm plot highlights the distribution of Shapley values, contrasting the predictive significance of pZVR and clinical variables between cases with GS = 3+3 or Grade Group = 1 (n = 23) and BPH‐Only (n = 106). Each point on the plot indicates how a particular variable contributes to the prediction for a single patient or sample. This visualization reveals the direction and magnitude of each variable's effect on the model's output, either augmenting or diminishing the predicted value relative to the mean.

FIGURE 10

Addition of pZVR to clinical parameter (age) improves distinguishing GGG = 1 prostate cancer from BPH‐Only compared to age alone. ROC curves from two models: TZ_PZ (pZVR) + age (AUC = 0.910) and age (AUC = 0.750).