Habitat analysis based on magnetic resonance imaging for the prediction of prostate cancer: a dual-center study

Abstract

Background

The application of habitat analysis is anticipated to enhance the diagnostic efficacy of magnetic resonance imaging (MRI) in prostate cancer (PCa) by providing a more accurate reflection of the microenvironmental characteristics within the lesion. The objective of this study was to investigate the feasibility of multisequence and multiregional MRI-based habitat analysis in the differentiation of PCa and benign prostatic hyperplasia (BPH).

Methods

We retrospectively evaluated the data of 673 cases from The Second Affiliated Hospital of Nanchang University and The First Hospital of Xiushui who received MRI examination of the prostate and pathologically confirmed diagnosis of PCa or BPH. Habitat features and classical radiomic features from the regions of lesions and prostate gland (PG) were extracted for model construction. Receiver operating characteristic analysis was used to assess the performance of the models. An integrated nomogram combining dominant models and clinical variables was ultimately constructed. In addition, we further assessed the performance of the nomogram in a subgroup of early-stage lesions without capsular invasion (CIV). The Delong test was used to compare the differences in the area under receiver operating characteristic curve (AUC) between models.

Results

The AUCs of the habitat radiomics score (rad-score) based on the lesion (LHrad-score) in both the internal (0.898) and external validation (0.878) sets were higher than those of the rad-score based on the lesion (0.860 and 0.854, respectively). The AUCs of the classical rad-score based on PG (PCrad-score; 0.883 and 0.865 in the internal and external sets, respectively) were higher than those of the habitat rad-score based on PG (0.871 and 0.773, respectively). By combining the PCrad-score and LHrad-score with clinically independent predictors, the nomogram yielded AUCs of 0.899 and 0.963 in the internal and external sets, respectively. Discrimination between early-stage PCa and BPH in the overall validation set yielded an AUC of 0.802.

Conclusions

The habitat analysis may serve as a means to noninvasively and preoperatively identifying PCa from BPH, even in the early stages of PCa.

Introduction

Prostate cancer (PCa) is a common disease among older men (1,2), and a variety of treatment options are available for managing this condition. However, benign prostatic hyperplasia (BPH) is also prevalent in aging males. Although therapeutic strategies for BPH exist, the treatment approaches for these two conditions differ significantly (3). Accurate diagnosis is therefore of critical importance, as it directly determines the choice of treatment and ultimately affects the patient’s prognosis.

Rectal palpation and prostate-specific antigen (PSA) are the primary methods in the clinical screening of PCa. However, they have limitations in terms of sensitivity and specificity and are susceptible to misdiagnosis and omission (4). Prostate biopsy, as the gold standard for diagnosing PCa, is not a suitable approach in PCa screening due to the potential risk of bleeding and infection and can even accelerate the spread of tumor (5). In order to reduce the false-positive rate and the number of unnecessary prostate biopsies, clinicians often combine multiparametric magnetic resonance imaging (MRI) for comprehensive evaluation, through which the detection rate of PCa can be enhanced (6,7). However, some lesions manifest ambiguous signal characteristics, which presents a challenge in differentiating between BPH and PCa, especially in the early stage of PCa without capsular invasion (CIV) (8,9). It is therefore necessary to develop objective quantitative methods that can assist in differentiating PCa from BPH.

Habitat analysis is an image-processing technique that divides a lesion into multiple subregions based on histopathological and molecular biological differences (10). This is achieved through unsupervised clustering based on the similarity of image features within a lesion. The similar voxels are grouped into the same tumor subregion, through which intralesionnal microenvironmental features can be obtained (11,12). Habitat analysis has been employed in the diagnosis, efficacy prediction, and prognostic evaluation of cancers in a multitude of locations throughout the body (13-15). Previous research has indicated that PCa often exhibits a complex intratumor microenvironment in comparison to other lesions (16,17). It is therefore possible that habitat analysis may enhance the ability to diagnose PCa via the analysis of the microenvironmental characteristics within the lesion.

This study aimed to evaluate the efficacy of habitat analysis in the diagnosis of PCa and to compare its performance with that of classical radiomics model. Furthermore, a combined model nomogram including clinical independent predictors was constructed to achieve an accurate diagnosis of PCa. We present this article in accordance with the CLEAR reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-223/rc).

Methods

This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the institutional and governmental ethics committees of The Second Affiliated Hospital of Nanchang University (No. IIT-0-2024-279) and The First Hospital of Xiushui (No. GSL 2025002). The requirement for individual consent was waived due to the retrospective nature of the analysis.

Patients

This study retrospectively recruited a total of 1,671 patients with PCa or BPH from two centers who had undergone prostate MRI and received a pathologically confirmed diagnosis of PCa between September 2017 and May 2023 at The Second Affiliated Hospital of Nanchang University (Hospital A) or between September 2020 and July 2024 at The First Hospital of Xiushui (Hospital B). The inclusion criteria were as follows: (I) prostate MRI examination; (II) complete clinical and imaging data; and (III) PCa or BPH pathologically confirmed within 1 month after MRI examination. Meanwhile, the exclusion criteria were as follows: (I) a history of prostate-related treatment (including chemotherapy, radiotherapy, or surgical treatment) before MRI examination; (II) an incomplete sequence or poor quality MRI; and (III) patients with a lesion size <3 in consecutive levels in MRI scans. The patients from Hospital A were randomly divided into a training set and an internal validation set at a ratio of 8:2. The external validation set comprised patients from Hospital B. The patient screening process is detailed in Figure 1.

Figure 1.

Patient selection flowchart. Hospital A, The Second Affiliated Hospital of Nanchang University; Hospital B, The First Hospital of Xiushui. BPH, benign prostatic hyperplasia; MRI, magnetic resonance imaging; PCa, prostate cancer.

Clinical data collection

The following clinical variables were collected from patients: age, PSA, prostate gland volume (PV), and PSA density (PSAD). The PV was defined with the following formula:

$$PV\left({\text{cm}}^3\right)=AP\left(\text{cm}\right)\times LD\left(\text{cm}\right)\times TD\left(\text{cm}\right)\times 0.52$$ [1]

where the AP and LD are the maximum anteroposterior dimension and the maximum longitudinal dimension, respectively, both of which were placed via midsagittal T2-weighted imaging (T2WI); and TD is the maximum transverse dimension (placed on an axial T2WI) (18). PSAD was calculated with the following formula:

$$PSAD=PSA\left(\text{ng/ml}\right)/PV\left({\text{cm}}^3\right)$$ [2]

Finally, clinically independent predictors were identified with a forward stepwise selection logistic regression and subsequently employed as clinical parameters for the construction of further models.

MRI acquisition and image processing

MR images from Hospital A were acquired with a 1.5-T MRI device (Signa HDxt 1.5 T, GE HealthCare, Chicago, IL, USA) and two 3.0-T MR devices (Signa HDxt 3.0 T and Discovery 3.0 T; GE HealthCare). MR images from Hospital B were acquired with a 3.0-T MR device (MAGNETOM Skyra 3.0 T, Siemens Healthineers, Erlangen, Germany). The scanning parameters of the two hospitals are presented in Table S1.

Prior to undertaking further analysis, all images were subjected to the following preprocessing steps. First, T2WI and T1-weighted imaging (T1WI) were aligned according to diffusion-weighted imaging (DWI) scans with a b-value of 0 mm²/s with the registration model in 3D Slicer (v. 5.6.1). Second, the voxels were resampled into 1 mm³ with the linear interpolation method. Finally, the pixels were truncated to the range of 0–4,300 to attenuate the effect of outliers.

Region of interest (ROI) segmentation

Two kinds of ROIs of lesion and the prostate gland (PG) for each patient were delineated by a radiologist. All ROIs were delineated on the T2WI sequence follows: (I) ROIs were delineated along the edge of the lesion or prostate as much as possible. (II) The urethra, necrosis, cysts, and urethral ducts were avoided. (III) If a patient had multiple lesions at the same time, the largest volume was obtained as the target lesion. (IV) Delineation was conducted layer by layer, the final volume of interest (VOI) was obtained via fusion of the ROIs of each layer, and then the VOI was duplicated to T1WI and DWI. Another radiologist randomly selected 50% of the cases (277 cases) in Hospital A for secondary delineation of the ROI outline.

Moreover, during the delineation process, an assessment of the CIV of lesions based on MRI was performed. The lesions with no breach of the prostate capsule, and no invasion of the periprostatic space or adjacent anatomical structures (e.g., seminal vesicles, rectum, and bladder) were classified into the early-stage lesion group (19).

Habitat analysis

Voxel vector features were acquired through the combination of three conventional MR sequences, including T1WI, T2WI, and DWI. An unsupervised cluster analysis based on the vector features was conducted with the k-means model. The squared Euclidean distance between voxels was employed as a similarity metric to evaluate voxel similarity within the VOI and to categorize analogous voxels into a subregion. The optimal number of clusters within a range from 2 to 5 was determined according to the Calinski-Harabasz score, with a larger score indicating a better performance (20).

Feature extraction

A total of 105 radiomics features were extracted based on the Pyradiomics package from each sequence of the VOI with a fixed bin width of 5. Before the extraction, image normalization was applied to the VOI in each image to enhance the comparability between images from different scanners. These features included 16 first-order features, 14 shape features, and 75 texture features, including gray-level co-occurrence matrix, gray-level run length matrix, gray-level size zone matrix, and neighborhood gray-tone difference matrix (21).

Classical radiomics features were extracted based on the overall VOI. The acquisition of habitat features was conducted as follows: (I) radiomics features were extracted from each subregion. (II) According to the varying number of clusters present in each patient, the k-nearest neighbor classification algorithm was employed to interpolate the habitat features that could be absent in order to retain the microenvironmental information. (III) The fusion of the subregion features was applied to derive the final habitat features.

Feature selection and model construction

All data of features were standardized with Z-scores to ensure that the parameters were of a similar magnitude and normally distributed. Feature reduction and model construction were performed in the training set. The feature reduction methods included the intraclass correlation coefficient, Pearson correlation analysis, and least absolute shrinkage and selection operator. The detailed process of features reduction is shown in the Appendix 1. The remaining radiomics features and their corresponding coefficients were combined with constants to obtain the radiomics formula, and the four following types of radiomics score (rad-score) were then acquired in each patient: the classical rad-score based on PG (PCrad-score), the classical rad-score based on the lesion (LCrad-score), the habitat rad-score based on the PG (PHrad-score), and the habitat rad-score based on the lesion (LHrad-score).

Finally, we identified the two best models by comparing the diagnostic efficacy of the models based on classical radiomics features with the models based on habitat features (LCrad-score vs. LH-rad-score; PCrad-score vs. PHrad-score). A nomogram was developed by combining the two selected models with clinically independent predictors. Figure 2 illustrates the flow of the study.

Figure 2.

The workflow of the habitat analysis in the study. CIV, capsular invasion; ICC, intraclass correlation coefficient; LassoCV, least absolute shrinkage and selection operator with Cross-Validation; LCrad-score, classical radiomics score based on the lesion; LHrad-score, habitat radiomics score based on the lesion; PCrad-score, classical radiomics score based on the prostate gland; PG, prostate gland; PHrad-score, habitat radiomics score based on the prostate gland; PSA, prostate-specific antigen; PSAD, prostate-specific antigen density.

Statistical analysis

All statistical analyses were conducted using the Python version 3.9.13 (Python Software Foundation, Wilmington, DE, USA) and R version 3.6.0 (The R Foundation for Statistical Computing, Vienna, Austria). A P value of less than 0.05 was considered to indicate statistical significance. The Kolmogorov-Smirnov test was used to assess the normality of the distribution. The comparison of continuous variables was performed with the two-sample t-test or the Mann-Whitney test, while the between-group differences for categorical variables were assessed with the Chi-squared test.

The interquartile range (IQR) method was employed for identifying outliers. The range of normal data was determined with the following formula:

$$\begin{array}{l}{Q}_R=Q_3-Q_1\\ x_{min}=Q_1-1.5Q_R\\ x_{max}=Q_3+1.5Q_R\end{array}$$ [3]

where the Q_R is the IQR of the given data set, Q₁ is the first quartile, Q₃ is the third quartile, and x_min and x_max are to the minimum and maximum values of the observed data, respectively. The data exceeding 1.5 times the IQR were classified as outliers. Prior to modelling, outliers were replaced by the mean of normal data to eliminate the impact of outliers.

Multivariate logistic regression was implemented for combined model construction. The performance of the model was comprehensively evaluated with five indicators: area under receiver operating characteristic curve (AUC), sensitivity, specificity, accuracy, and F1 value. The Delong test was used to compare the AUCs between models. The Hosmer-Lemeshow test was used to assess the fit of the model to the pathological results, and decision curve analysis (DCA) was used to evaluate the clinical benefits of the models.

Results

Patient characteristics

A total of 673 patients from two hospitals were included in this study, and the clinical information of the patients is presented in Table 1. From Hospital A, there were 204 cases of BPH and 381 cases of PCa, while from Hospital B, there were 60 cases of BPH and 28 cases of PCa. The patient information for the training set (468 patients), internal validation set (117 patients), and external validation set (88 patients) is presented in Table 2. The forward-stepwise selection logistic regression analysis demonstrated that PV and PSAD were independent predictors of PCa (P<0.001).

Table 1. Clinical characteristics of patients from the two hospitals.

Characteristic	Hospital A	Hospital B	P
Pathological result			<0.001
BPH	204	60
PCa	381	28
Age (years)	71.174±7.582	71.045±8.161	0.883
PSA (ng/mL)	19.600 (9.720, 51.940)	8.445 (3.690, 14.468)	<0.001
PV (cm3)	43.010 (30.010, 62.170)	39.010 (26.848, 66.412)	0.171
PSAD	0.490 (0.190, 1.260)	0.240 (0.087, 0.823)	<0.001

Data are presented as number, mean ± standard deviation, or median (interquartile range). Hospital A, The Second Affiliated Hospital of Nanchang University; Hospital B, The First Hospital of Xiushui. BPH, benign prostatic hyperplasia; PCa, prostate cancer; PSA, prostate-specific antigen; PSAD, prostate-specific antigen density; PV, prostate gland volume.

Table 2. Clinical characteristics of the different sets.

Characteristic	Training set (n=468)	Internal validation set (n=117)	External validation set (n=88)
Pathological result
BPH	169 [36]	35 [30]	60 [68]
PCa	299 [64]	82 [70]	28 [32]
Age (years)	71.096±7.587	71.487±7.585	71.045±8.161
PSA (ng/mL)	19.100 (9.670, 49.233)	20.900 (10.800, 54.860)	8.445 (3.690, 14.468)
PV (cm3)	42.155 (29.460, 59.860)	46.290 (33.170, 71.370)	39.010 (26.848, 66.412)
PSAD	0.485 (0.190, 1.272)	0.490 (0.220, 1.200)	0.240 (0.087, 0.823)

Data are presented as number [%], mean ± standard deviation, or median (interquartile range). BPH, benign prostatic hyperplasia; PCa, prostate cancer; PSA, prostate-specific antigen; PSAD, prostate-specific antigen density; PV, prostate gland volume.

Habitat imaging

For the habitat analysis based on PG, the optimal number of clusters was two, with a maximum of two clusters permitted. According to the clustering results, 673 patients had subregion 1 features and 541 patients had subregion 2 features. Figure 3A,3B illustrates the distribution of each subregion. The number of patients with subregion 1 in the BPH group was comparable to that of the PCa group, while the number of patients with subregion 2 was lower than that of the PCa group.

Figure 3.

The clustering results from habitat analysis. (A,C) The proportion of each subregion in BPH or PCa. (B,D) The percentage of BPH and PCa in each subregion partition. BPH, benign prostatic hyperplasia; PCa, prostate cancer.

For the habitat analysis based on lesions, the optimal number of clusters was three, with a maximum of three clusters permitted. The clustering results demonstrated that 493 patients had subregion 1, 1,661 patients had subregion 2, and 251 patients had subregion 3. Figure 3C,3D illustrates the distribution of each subregion; the BPH group had a lower number of subregion 1 features and subregion 3 than features did the PCa group, and both groups had approximately similar proportions of subregion 2 features. Figure S1 showed a representative case of habitat imaging.

Model construction and evaluation

A total of 105 radiomics features were extracted for each sequence/subregion. After the feature reduction, 27 classical radiomics features from PG, 35 classical radiomics features from lesions, 57 habitat features from PG, and 30 habitat features from lesions were retained. The detailed feature selection process and the radiomics formula are presented in Appendix 2 and Figure S2. Based on the corresponding radiomics formula, four radiomics models were constructed: PCrad-score, PHrad-score, LCrad-score, and LHrad-score. The contributions of different sequences in the models are shown in Table S2. The results demonstrate that in the model comprising the PCrad-score and LCrad-score, the hierarchical contribution of sequences was as follows: DWI > T2WI > T1WI; meanwhile, in the PHrad-score and LHrad-score, the distribution was as follows: T2WI > DWI > T1WI.

Table 3 shows the performance of the models. The LHrad-score demonstrated the highest efficacy, with AUCs of 0.935 [95% confidence interval (CI): 0.913–0.957], 0.898 (95% CI: 0.841–0.955), and 0.878 (95% CI: 0.802–0.954) in the training set, internal validation set, and external validation set, respectively. The specificity in the external validation set was 93.3%. Furthermore, the AUC of the LHrad-score was higher than that of LCrad-score in the internal set (0.860 vs. 0.898; P=0.396) and external validation set (0.854 vs. 0.878; P=0.040). The AUC of PCrad-score was higher than that of PHrad-score in both the internal set (AUC 0.883 vs. 0.871; P=0.713) and external validation set (AUC 0.865 vs. 0.773; P=0.447), but the differences were not statistically significant. Consequently, the LHrad-score and PCrad-score, given their superior efficacy, were selected for inclusion in the construction of the combined model.

Table 3. The comparison of the performance between the habitat models and clinical radiomics models.

Model	AUC (95% CI)	ACC	SEN	SPE	F1	P
Training set
PCrad-score	0.941 (0.920–0.963)	0.876	0.873	0.882	0.900	<0.001
PHrad-score	0.965 (0.950–0.979)	0.906	0.910	0.899	0.925	<0.001
LCrad-score	0.952 (0.934–0.971)	0.895	0.896	0.893	0.916	<0.001
LHrad-score	0.935 (0.913–0.957)	0.846	0.793	0.941	0.869	<0.001
Internal validation set
PCrad-score	0.883 (0.814–0.951)	0.812	0.756	0.943	0.849	<0.001
PHrad-score	0.871 (0.791–0.951)	0.863	0.878	0.829	0.900	<0.001
LCrad-score	0.860 (0.766–0.953)	0.897	0.927	0.829	0.927	<0.001
LHrad-score	0.898 (0.841–0.955)	0.829	0.780	0.943	0.865	<0.001
External validation set
PCrad-score	0.865 (0.779–0.951)	0.830	0.857	0.817	0.762	<0.001
PHrad-score	0.773 (0.661–0.885)	0.716	0.821	0.667	0.648	<0.001
LCrad-score	0.854 (0.765–0.943)	0.818	0.750	0.850	0.724	<0.001
LHrad-score	0.878 (0.802–0.954)	0.852	0.679	0.933	0.745	<0.001

ACC, accuracy; AUC, area under receiver operating characteristic curve; CI, confidence interval; LCrad-score, classical radiomics score based on the lesion; LHrad-score, habitat radiomics score based on the lesion; PHrad-score, habitat radiomics score based on the prostate gland; PCrad-score, classical radiomics score based on the prostate gland; SEN, sensitivity; SPE, specificity.

Combined model and nomogram

A logistic regression-based combined model was constructed through a combination of the PCrad-score, LHrad-score, and clinically independent predictors, and the corresponding nomogram was constructed (Figure 4). Figure 5 illustrates receiver operating characteristic curves, calibration curves, and decision curves of the models, while Table 4 presents the performance of the models. In comparison to the other models, the combined model yielded the best performance in the training set, internal validation set, and external validation set with AUCs of 0.972 (95% CI: 0.960–0.984), 0.899 (95% CI: 0.830–0.968), and 0.963 (95% CI: 0.929–0.996), respectively. The results of Delong test between models are shown in the Figure S3. The calibration curves indicated that the consistency between the combined model and the actual results was good (Hosmer-Lemeshow test: all P values >0.05). The decision curve suggests that the clinical benefit of the combined model was higher than that of the other models.

Figure 4.

Nomogram for differentiating between PCa and BPH. BPH, benign prostatic hyperplasia; LHrad-score, habitat radiomics score based on the lesion; PCa, prostate cancer; PCrad-score, classical radiomics score based on the prostate gland; PSAD, prostate-specific antigen density; PV, prostate gland volume.

Figure 5.

Performance of models in differentiating between PCa and BPH. (A-C) Receiver operating characteristic curves, (D-F) calibration curves, and (G-I) decision curves. BPH, benign prostatic hyperplasia; CI, confidence interval; LHrad-score, the model based on habitat rad-score based on lesion; PCrad-score, classical rad-score based on prostate gland; PCa, prostate cancer.

Table 4. The comparison of performance between the combined models and the clinical models.

Model	AUC (95% CI)	ACC	SEN	SPE	F1	P
Training set
Clinical model	0.866 (0.833–0.899)	0.803	0.816	0.781	0.841	<0.001
Clinical + PCrad-score	0.955 (0.938–0.972)	0.897	0.883	0.923	0.917	<0.001
Clinical + LHrad-score	0.960 (0.944–0.975)	0.904	0.910	0.893	0.924	<0.001
Combined model	0.972 (0.96–0.984)	0.921	0.926	0.911	0.972	<0.001
Internal validation set
Clinical model	0.768 (0.655–0.881)	0.795	0.817	0.743	0.848	<0.001
Clinical + PCrad-score	0.882 (0.809–0.954)	0.829	0.805	0.886	0.868	<0.001
Clinical + LHrad-score	0.877 (0.804–0.95)	0.846	0.829	0.886	0.883	<0.001
Combined model	0.899 (0.83–0.968)	0.863	0.829	0.943	0.895	<0.001
External validation set
Clinical model	0.826 (0.723–0.928)	0.818	0.714	0.867	0.714	<0.001
Clinical + PCrad-score	0.889 (0.808–0.969)	0.841	0.893	0.817	0.781	<0.001
Clinical + LHrad-score	0.951 (0.910–0.993)	0.886	0.964	0.85	0.844	<0.001
Combined model	0.963 (0.929–0.996)	0.875	0.964	0.833	0.831	<0.001

ACC, accuracy; AUC, area under receiver operating characteristic curve; CI, confidence interval; LHrad-score, the model based on habitat rad-score based on lesion; PCrad-score, classical rad-score based on prostate gland; SEN, sensitivity; SPE, specificity.

Performance of models in the noncapsular invasion subgroup

It is of paramount importance to accurately diagnose PCa at its early stages without CIV to prevent its progression to CIV, which in turn improves patient prognosis (22). However, it is difficult to distinguish early-stage PCa from BPH. Consequently, we proceeded to conduct a further analysis of the diagnostic efficacy of each model for the early noncapsular invasion subgroup in the training set and total validation set (comprising the internal validation set and the external validation set). The training set comprised 167 benign and 130 PCa cases, while the validation set comprised 94 benign and 43 PCa cases.

The performance of the models is presented in the Appendix 3, Table S3 and Figure S4. The results demonstrated that the AUC of the training set for the combined model was 0.952 (95% CI: 0.930–0.974), which was markedly higher than that of the clinical model (AUC =0.813; 95% CI: 0.764–0.862; P<0.001). The AUC of the combined model in the validation set was marginally higher than that of the clinical model, but the difference was not statistically significant (0.802 vs. 0.764; P=0.454). The results of Delong test are shown in Figure S5. Furthermore, the calibration curves demonstrated that there was no observable difference between the combined model and the actual results (Hosmer-Lemeshow test: both P values >0.05). The decision curves indicated that the clinical benefit of the combined model was superior to that of the other models.

Discussion

In this study, four radiomics models were constructed based on classical radiomics or habitat analysis, including the PCrad-score, LCrad-score, PHrad-score, and LHrad-score. The LHrad-score demonstrated the most favorable performance, with an AUC of 0.898 (95% CI: 0.841–0.955) in the internal validation set and 0.878 (95% CI: 0.802–0.954) in the external validation set. The nomogram, which combined the LHrad-score, PCrad-score, and clinically independent predictors, yielded a higher performance than did the other models. In a subgroup of lesions without CIV, the combined model maintained a superior performance in the diagnosis of PCa.

Given that both PCa and BPH can be multiple and invade the surrounding PG, it is possible that the rest of the PG beyond the most prominent lesion may also contain information related to heterogeneity. Accordingly, we delineated both the PG and the lesion. Previous studies have documented that interscanner variability and multicenter imaging protocols may compromise feature robustness (23,24). In this study, the images were acquired by four scanners from two hospitals. To mitigate the effects, standardized preprocessing involving image normalization and fixed bin-width discretization was implemented prior to feature extraction. Features demonstrating high stability (intraclass correlation coefficient >0.75) were retained for modeling. The results revealed that conventional radiomics models achieved good performance across both the internal and external validation sets, with AUCs of 0.854–0.883, confirming the effectiveness of the imaging preprocess. The classical radiomics models based on PGs exhibited slightly greater performance in terms of AUCs than did those based on the lesions (internal validation set: 0.883 vs. 0.860; external validation set: 0.865 vs. 0.854). This is in line with Han et al.’s study, in which the AUC of models based on PG and on lesions were 0.731 and 0.710, respectively (25). This general finding may be attributed to the inherent lack of signal specificity in extralesional regions, from which supplementary data fail to provide a significant improvement.

In recent years, there have been numerous studies employing habitat analysis for the diagnosis and assessment of tumors. Huang et al. used habitat analysis based on computed tomography images and clinical factors, successfully predicting the response to radiofrequency ablation of lung metastases from colon cancer (26). Shi et al. developed a quantitative index based on habitat analysis, which successfully predicted the pathological complete response in patients with breast cancer following neoadjuvant chemotherapy (13). However, it remains unclear whether this technique can improve the diagnosis and assessment of PCa. In this study, we performed a habitat analysis and extracted the corresponding habitat features based on the PG and lesions. For the models based on lesions, the AUC of the LHrad-score was higher than that of the classical radiomics model (internal validation set: 0.898 vs. 0.860; external validation set: 0.878 vs. 0.854). The optimal clustering for habitat imaging in the lesion region was identified as 3. The largest number of patients had subregion 2, and the proportion of PCa and BPH in subregion 2 was comparable. PCa occupied a significantly higher proportion of subregion 3 and subregion 1 than BPH. This suggests that PCa tends to have a high degree of intralesion heterogeneity. After dimension reduction, features from subregion 2 had the largest proportion and the highest weight. This may be attributable to the greater number of subregion 2 features initially obtained and their more extensive distribution in both lesions. For the models based on the PG, the PCrad-score exhibited a higher AUC than did the PHrad-score (internal validation set: 0.883 vs. 0.871; external validation set: 0.865 vs. 0.773). This finding is inconsistent with the initial hypothesis of this study. Upon examination of the habitat subregion segmentation, it was observed that a comparable number of patients with subregion 1 and subregion 2 were present in PCa and BPH. This can be explained by the high proportion of normal regions present in each PG and the evident signal differentiation between the normal and nonnormal regions. This could render habitat analysis being less effective in the subregion segmentation of PG and susceptible to the redundant features of the normal prostate background, resulting in a reduction of model performance.

In this study, PV, PSAD, PCrad-score and LHrad-score were included as independent variables in the construction of the clinical-imaging nomogram. For the differentiating of PCa and BPH in the internal and external validation sets, the AUCs were 0.899 and 0.963, respectively. These results are in line with the previous studies, while the AUC in the external validation set is slightly higher (27,28). This could be explained that the habitat analysis can provide more information about the environment within a lesion than can classical radiomics. Furthermore, it is challenging to distinguish PCa from BPH in the lesions without CIV. However, when PCa progresses to peritoneal invasion, it significantly impacts patient prognosis (22). This study further examined the performance of each model in this type of lesion. The results indicated that the AUC of all models was lower. The combined model yielded an AUC of 0.802 in the validation set; a sensitivity and specificity of 0.674 and 0.894, respectively; and good migration performance. In the of guidance clinical biopsy, the combined model demonstrated superior performance as compared to the clinical models, avoiding 83.3–94.3% unnecessary biopsies (specificity) at the cost of 3.6–17.1% missed PCa cases (1 − sensitivity). In summary, the combined model provides enhanced clinical net benefit and enables accurate diagnosis of lesions without CIV. It can facilitate earlier diagnosis and ultimately improve patient prognosis.

Limitations

This study involved certain limitations which should be addressed. First, we employed a retrospective analysis, and no prospective experiment was employed to validate the model efficacy. Further investigations with larger sample sizes and prospective designs are warranted to further validate the model’s clinical efficacy. Second, Hospital B, which was used as the validation set, exhibited significant differences in PSA, PSAD, and the composition of the included cases in terms of both benign and malignant elements, as compared to Hospital A. Only age and PV were similar between the two hospitals, which may be due to the differences in the patient populations admitted to these centers. Third, lesion and PG region delineation in this study was performed manually, constituting a labor-intensive process. This procedure is required for the implementation of artificial intelligence-driven automated segmentation to improve operational efficiency and reduce labor. Finally, the value of apparent diffusion coefficient maps and raw images in the enhancement of model performance needs to be further verified in future studies.

Conclusions

To the best of our knowledge, this is the first study to examine the application of habitat analysis in the prediction of PCa. The findings in this study indicate that habitat analysis has the potential for the noninvasive and preoperative identification of PCa from BPH, even in the early stages of PCa.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the institutional and governmental ethics committees of The Second Affiliated Hospital of Nanchang University (No. IIT-0-2024-279) and The First Hospital of Xiushui (No. GSL 2025002). The requirement for individual consent was waived due to the retrospective nature of the analysis.

Data Sharing Statement

Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-223/dss

qims-15-09-8395-dss.pdf ^{(120.5KB, pdf)}

DOI: 10.21037/qims-2025-223