Multimodal imaging deep learning model for predicting extraprostatic extension in prostate cancer using MpMRI and 18 F-PSMA-PET/CT

Abstract

Objective

This study aimed to construct a multimodal imaging deep learning (DL) model integrating mpMRI and ¹⁸F-PSMA-PET/CT for the prediction of extraprostatic extension (EPE) in prostate cancer, and to assess its effectiveness in enhancing the diagnostic accuracy of radiologists.

Methods

Clinical and imaging data were retrospectively collected from patients with pathologically confirmed prostate cancer (PCa) who underwent radical prostatectomy (RP). Data were collected from a primary institution (Center 1, n = 197) between January 2019 and June 2022 and an external institution (Center 2, n = 36) between July 2021 and November 2022. A multimodal DL model incorporating mpMRI and ¹⁸F-PSMA-PET/CT was developed to support radiologists in assessing EPE using the EPE-grade scoring system. The predictive performance of the DL model was compared with that of single-modality models, as well as with radiologist assessments with and without model assistance. Clinical net benefit of the model was also assessed.

Results

For patients in Center 1, the area under the curve (AUC) for predicting EPE was 0.76 (0.72–0.80), 0.77 (0.70–0.82), and 0.82 (0.78–0.87) for the mpMRI-based DL model, PET/CT-based DL model, and the combined mpMRI + PET/CT multimodal DL model, respectively. In the external test set (Center 2), the AUCs for these models were 0.75 (0.60–0.88), 0.77 (0.72–0.88), and 0.81 (0.63–0.97), respectively. The multimodal DL model demonstrated superior predictive accuracy compared to single-modality models in both internal and external validations. The deep learning-assisted EPE-grade scoring model significantly improved AUC and sensitivity compared to radiologist EPE-grade scoring alone (P < 0.05), with a modest reduction in specificity. Additionally, the deep learning-assisted scoring model provided greater clinical net benefit than the radiologist EPE-grade score used by radiologists alone.

Conclusion

The multimodal imaging deep learning model, integrating mpMRI and 18 F-PSMA PET/CT, demonstrates promising predictive performance for EPE in prostate cancer and enhances the accuracy of radiologists in EPE assessment. The model holds potential as a supportive tool for more individualized and precise therapeutic decision-making.

Introduction

Prostate cancer (PCa) is one of the most prevalent malignant tumors affecting the male genitourinary system. According to the Global Cancer Statistics 2020 report by the World Health Organization’s International Agency for Research on Cancer in February 2021, prostate cancer accounted for 7.3% of all new cancer cases worldwide in 2020, ranking fourth in incidence and eighth in mortality [1]. PCa has become a significant health concern for men in China, with extraprostatic extension (EPE) observed in approximately 20–30% of cases. Accurate assessment of EPE is essential not only for surgical planning but also for focal therapy eligibility, radiation planning and other factors in patients with PCa. For EPE-negative patients, preserving the adjacent neurovascular bundle during surgery can lower the risk of postoperative complications, such as erectile dysfunction and urinary incontinence, while maintaining tumor control. Conversely, in patients with EPE, complete resection of the prostate and surrounding neurovascular bundle is necessary to minimize the risk of poor outcomes, including positive surgical margins, biochemical recurrence, and distant metastasis [2, 3]. Thus, accurate EPE assessment is crucial for optimal surgical planning, postoperative quality of life, and tumor prognosis.

Deep learning is an artificial intelligence approach that can automatically learn feature representations from data by utilizing artificial neural networks. It can automatically learn specific features necessary for task completion which makes it possible to establish more abstract features, resulting in greater informational content and enhanced generalizability [4]. Recent research highlights the potential of deep learning (DL) models in predicting EPE. Notably, to the best of our knowledge, few studies have yet reported on developing DL models for EPE prediction using multimodal imaging using both mpMRI and PSMA-PET/CT. Most of the studies until now assessing EPE deal with single modalities (MRI or PSMA-PET/CT) alone or comparison between them, only few studies have aimed to conduct and integrated multimodality evaluation. The objective of this study is to construct an EPE prediction model using residual convolutional neural networks based on mpMRI and ¹⁸F-PSMA-PET/CT multimodal imaging and to assess its potential utility as a supportive tool for radiologists in preoperative EPE assessment.

Materials and methods

Study population

This retrospective study was conducted with the approval of the Ethics Committee of The First Affiliated Hospital of Wenzhou Medical University (Approval Number: KY2024-R181). The included patients diagnosed with PCa who underwent radical prostatectomy (RP) between January 2019 and June 2022 at the primary institution (Center 1) and between July 2021 and November 2022 at the external institution (Center 2). Participants were required to have complete clinical and imaging data. Exclusion criteria were as follows: (1) absence of preoperative mpMRI and 18 F-PSMA-PET/CT imaging data within a four-week period; (2) administration of neoadjuvant therapy prior to imaging or surgery; (3) incomplete clinical or pathological data; and (4) diagnosis of other malignant tumors alongside PCa. A total of 197 cases from Center 1 were included and randomly assigned to training (n = 177) and testing (n = 20) groups using 10-fold cross-validation, while 36 cases from Center 2 comprised the external test group (n = 36) (Fig. 1).

Fig. 1.

Patient enrollment flowchart

MRI and PET/CT image acquisition

MRI and PSMA-PET/CT imaging protocols at the center adhered to those outlined in Part Two. At the external center, multiparametric MRI (mpMRI) scans were conducted using either the GE (Signa Architect) or United Imaging (uMR 880) 3.0T MR scanners, following the guidelines set forth in the Consensus on PCa MRI Examination and Diagnosis (Second Edition). Axial T2-weighted imaging (T2WI) was acquired using a 2D fast spin echo sequence, with or without fat suppression, with the following parameters: repetition time (TR) of 3000 ms, echo time (TE) of 90–110 ms, echo train length of 20–30, and a slice thickness of 3 mm. The field of view (FOV) was approximately 250 mm × 250 mm, covering the entire prostate and bilateral seminal vesicles. Diffusion-weighted imaging (DWI) was carried out using single-shot spin-echo echo-planar imaging (SE-EPI) with b-values ≥ 800 s/mm², and apparent diffusion coefficient (ADC) maps were generated using formula (1).

Dynamic contrast-enhanced (DCE) imaging was performed using three-dimensional T1-weighted (T1WI) gradient-echo sequences, with a TR of less than 100 ms, TE of less than 5 ms, and a slice thickness of 3 mm. A gadolinium-based contrast agent (GBCA) was administered via cubital vein bolus injection at a dose of 0.1–0.2 mmol/kg, with an injection rate of 2–3 ml/s, followed by an equal volume of normal saline prior to DCE scanning.

At the external center, PSMA-PET/CT images were obtained using a Siemens PET/CT scanner (Biograph Vision; Siemens, Germany). Approximately 120 min following intravenous injection of ¹⁸F-PSMA-1007 at 4.0 MBq/kg, a low-dose CT plain scan was conducted from the skull base to mid-thigh with the following parameters: tube voltage of 130 KV, tube current of 110 mA, collimation of 64 mm × 0.625 mm, pitch of 0.85, rotation speed of 0.5 s, scanning slice thickness of 5 mm, and a reconstruction interval of 2.5 mm. The slice thickness for PET scanning was matched to that of the CT scan. PET images were acquired with a field of view of 570 mm, a matrix of 144 × 144, and a scanning slice thickness and interval of 5 mm using a three-dimensional model. Emission scanning was conducted with a time of 1.5 min per bed position, with a 50% overlap between adjacent bed positions. CT attenuation-corrected PET images were reconstructed using the time-of-flight algorithm.

Image preprocessing, DL model construction, and evaluation

To enhance focus on the prostate region, all patient images were first cropped around the region of interest (ROI) to isolate the prostate area, following the delineation as showed in Fig. 2. Images were then resampled to uniform dimensions (64 × 64 × 8), where 64 × 64 represents the image width and height, and 8 represents the number of slices. Subsequently, image normalization was applied to scale pixel values to a range of 0 to 1, with the objective to accelerate convergence of the deep learning network. For data augmentation, random affine transformations were used to reduce model overfitting and enhance generalization ability.

Fig. 2.

Deep Residual Convolutional Neural Network (ResNet-50) architecture diagram

A deep residual learning convolutional neural network (CNN), specifically ResNet-50, was used to develop the DL model for predicting EPE. The process involved the following steps: (1) Data preprocessing, where input images were converted into feature maps using a convolutional layer followed by a max pooling layer; (2) Feature extraction via concatenated bottleneck structures, using the ReLU activation function to retain non-linear features, and further enhanced through multiple residual blocks to deepen the network; and (3) Transformation of extracted features into class probabilities through an average pooling layer and fully connected layer, with the Softmax activation function applied to the output to yield a probability distribution, thus facilitating final classification (see Fig. 2).

For multimodal image fusion, an input-level fusion strategy was implemented by concatenating five different modality images (ADC, DWI, T2WI, CT, and PET) into multi-channel images to serve as network input. This approach enabled the retention of original image information, facilitating enhanced feature learning. During the model training phase, a ten-fold cross-validation technique was used to divide the data into training and testing sets, minimizing the risk of overfitting. The Cross Entropy Loss function was used for training, while the Adam optimizer was applied with a learning rate of 0.00001, a batch size of 16, and 300 iterations.

The performance of the model was assessed by calculating sensitivity, specificity, positive predictive value, negative predictive value, AUC, and clinical net benefit.

Construction of DL-assisted physician EPE-grade score model

Initially, a radiologist with over 10 years of experience in genitourinary imaging diagnostics, blinded to pathological results, assessed EPE risk based on the multiparametric MRI (mpMRI) images of patients using the EPE-grade scoring system proposed by Mehralivand et al. [5] This scoring system included the following criteria: 0 points for no MRI signs related to EPE; 1 point for a tumor capsular contact length greater than 1.5 cm or an irregular capsule and/or bulging; 2 points for both a tumor capsular contact length > 1.5 cm and an irregular capsule and/or bulging; and 3 points for obvious signs of EPE.

The DL model predictions were incorporated to adjust the radiologist’s EPE-grade scores according to specific rules: if the DL model prediction indicated a positive EPE, the radiologist’s score was increased by 1 point, except when the original score was 3 points, in which case it remained unchanged. Conversely, if the DL model prediction was negative, the radiologist’s score was reduced by 1 point, except when the original score was 0 points, which also remained unchanged. This methodology was used to develop the deep learning-assisted radiologist EPE-grade score model and to assess its diagnostic performance.

Pathological analysis

A pathologist with over 5 years of experience in urological pathology, blinded to clinical imaging results, reviewed and confirmed the postoperative pathological findings using whole-mount specimens sectioned at 3–4 mm intervals from apex to base. This assessment followed the prostate cancer pathological diagnostic criteria established by the International Society of Urological Pathology (ISUP) in 2014. These findings served as the final pathological reference standard for the study.

Statistical analysis

Statistical analysis was conducted using IBM SPSS software (V25.0; SPSS) and R software (V4.0.2). For continuous variables with a normal distribution, comparisons were conducted using the t-test; for non-normally distributed continuous variables, the rank sum test was applied. Categorical variables were compared using the χ² test or Fisher’s exact test. Model performance was assessed by analyzing receiver operating characteristic (ROC) curves and metrics derived from confusion matrixes. Differences between areas under different ROC curves were compared using the DeLong test. Clinical use of the model was assessed through decision curve analysis (DCA), where net benefit values were calculated; higher net benefit at specified risk thresholds indicated greater clinical value. P < 0.05 was considered statistically significant.

Results

General clinical data

A total of 197 and 36 patients with PCa meeting the inclusion criteria were enrolled from Centers 1 and 2, respectively. In Center 1, there were 107 (54.30%) EPE-positive and 90 (45.70%) EPE-negative cases. The mean age was 67.88 years, with a mean BMI of 24.11, median t-PSA of 10.72 ng/mL, and mean prostate volume of 34.41 ml. Statistically significant differences were identified (P < 0.05) between the EPE-negative and EPE-positive groups in terms of preoperative t-PSA, f-PSA, f/t-PSA ratio, PI-RADS score, EPE-grade classification, ISUP-GS grade, seminal vesicle invasion, and rate of positive surgical margins. No significant differences (P > 0.05) were found between groups for age, BMI, and prostate volume.

In Center 2, there were 18 (50.00%) EPE-positive (50.00%) and 18 EPE-negative cases. The mean age was 68.28 years, with a mean BMI of 23.44, median t-PSA of 16.05 ng/mL, and mean prostate volume of 27.69 ml. Statistically significant differences (P < 0.05) were observed between EPE-negative and EPE-positive groups in preoperative t-PSA, PI-RADS score, and ISUP-GS grade. No significant differences (P > 0.05) were found between groups for age, BMI, f-PSA, f/t-PSA ratio, prostate volume, EPE-grade classification, seminal vesicle invasion, or positive surgical margin rate. For further details, see Table 1.

Table 1.

Demographic information of patients

	Total (n = 233)	Center 1 (n = 197)	Center 2 (n = 36)	P
Age (years)	67.94 ± 6.37	67.88 ± 6.67	68.28 ± 4.50	0.654
Body mass index (BMI)	24.00 ± 2.64	24.11 ± 2.65	23.44 ± 2.56	0.165
t-PSA (ng/mL)	10.72 (7.82, 20.30)	10.72 (7.77, 17.38)	16.05 (8.94, 36.44)	0.013
f-PSA (ng/mL)	1.25 (0.91, 2.09)	1.25 (0.86, 1.99)	1.66 (1.05, 3.58)	0.055
f/t-PSA	0.10 (0.08, 0.14)	0.10 (0.09, 0.14)	0.10 (0.08, 0.14)	0.484
Prostate volume (ml)	32.32 (24.53, 43.50)	34.41 (25.76, 44.58)	27.69 (23.27, 35.54)	0.047
EPE-grade Score, n (%)				0.076
0	125 (53.65)	106 (53.81)	19 (52.78)
1	57 (24.46)	47 (23.86)	10 (27.78)
2	41 (17.60)	38 (19.29)	3 (8.33)
3	10 (4.29)	6 (3.05)	4 (11.11)
PIRADS, n (%)				0.450
1–2	3 (1.29)	3 (1.52)	0 (0.00)
3	50 (21.46)	43 (21.83)	7 (19.44)
4	54 (23.18)	42 (21.32)	12 (33.33)
5	126 (54.08)	109 (55.33)	17 (47.22)
ISUP-GS Score, n (%)				0.267
1	12 (5.15)	12 (6.09)	0 (0.00)
2	71 (30.47)	62 (31.47)	9 (25.00)
3	83 (35.62)	68 (34.52)	15 (41.67)
4	28 (12.02)	21 (10.66)	7 (19.44)
5	39 (16.74)	34 (17.26)	5 (13.89)
EPE, n (%)				0.633
Positive	125 (53.65)	107 (54.30)	18 (50.00)
Negative	108 (46.35)	90 (45.70)	18 (50.00)
SVI, n (%)				0.580
Positive	38 (16.31)	31 (15.74)	7 (19.44)
Negative	195 (83.69)	166 (84.26)	29 (80.56)
Surgical Margin, n (%)				0.005
Positive	87 (37.34)	66 (33.50)	21 (58.33)
Negative	146 (62.66)	131 (66.50)	15 (41.67)

Performance of DL models based on different imaging modalities

In Center 1, the AUC for EPE prediction was 0.76 (95% CI: 0.72–0.80) for the mpMRI-DL model, 0.77 for the PET/CT-DL model, and 0.82 for the combined mpMRI and PET/CT multimodal imaging DL model. In the external test group (Center 2), the AUCs for these three models were 0.75, 0.77, and 0.81, respectively. The multimodal imaging DL model demonstrated superior predictive performance compared to the single-modality DL models in both internal and external validations. For further details see Table 2.

Table 2.

Performance of DL models using different imaging modalities

Model	AUC (95% CI)	SEN	SPE	NPV	PPV	ACC
Center 1
mpMRI	0.76 (0.72–0.80)	0.65	0.75	0.74	0.73	0.72
PET/CT	0.77 (0.70–0.82)	0.76	0.71	0.79	0.69	0.74
Multimodal	0.82 (0.78–0.87)	0.79	0.78	0.81	0.75	0.78
Center 2
mpMRI	0.75(0.60–0.88)	0.67	0.53	0.89	0.44	0.69
PET/CT	0.77(0.72–0.88)	0.83	0.38	0.57	0.58	0.67
Multimodal	0.81(0.63–0.97)	0.94	0.56	0.93	0.70	0.78

Performance of DL-assisted EPE-grade score model

In Center 1, the AUC for preoperative EPE prediction based on the EPE-grade score of the radiologists from mpMRI image features was 0.64, while the DL-assisted EPE-grade score model achieved an AUC of 0.82. In the external test group (Center 2), the AUCs for these models were 0.68 and 0.87, respectively. The DL-assisted EPE-grade score model revealed a significantly improved AUC and sensitivity compared to the EPE-grade score of the radiologist alone (P < 0.05), with a slight decrease in specificity. For further details, see Table 3; Fig. 3.

Table 3.

Comparison of EPE-grade score and DL-assisted EPE-grade score model performance

Model	SEN	SPE	PPV	NPV	AUC	P
Center 1
EPE-Grade	0.34	0.88	0.70	0.61	0.64	< 0.001
DL/EPE-Grade	0.82	0.69	0.69	0.82	0.82
Center 2
DL/EPE-Grade	0.61	0.67	0.65	0.67	0.68	0.001
EPE-Grade	1.00	0.56	0.69	1.00	0.87

Fig. 3.

ROC curves for the three models in Center 1 (A) and Center 2 (B)

Clinical decision curve analysis of three models

The DCA curve (Fig. 4) indicates that when the EPE probability exceeds 0.6, the clinical net benefit of the DL-assisted EPE-grade score model is significantly higher than that of both the radiologist EPE-grade score alone and the DL model. This indicates that the DL-assisted radiologist EPE-grade score model provides a greater clinical net benefit.

Fig. 4.

Decision curve analysis of the three models. The threshold probability is shown by the x-axis, while the y-axis displays the net benefit. Notes: All: intervention for all patients; None: indicates no intervention for any patient

Discussion

Deep learning facilitates the processing and transformation of data through multilayer neural networks that emulate the functionality of neuronal structures, progressively extracting and abstracting features from images. The incorporation of residual structures in both the encoding and decoding convolutional layers, known as residual convolutional neural network (Res-Nets) enhances model performance [6]. In this study, a 50-layer residual convolutional neural network (ResNet-50) was used to develop a model for predicting EPE using multimodal imaging from mpMRI and PSMA-PET/CT scans. The results indicated that the multimodal imaging DL model outperformed single-modality models, achieving AUC values of 0.82 and 0.81 in Centers 1 and 2, respectively, thus validating its predictive efficacy with external data.

In recent years, applications of artificial intelligence (AI) in medicine have significantly expanded, demonstrating substantial potential across various domains, including assisted image recognition and segmentation, intelligent disease diagnosis and staging, and prognosis assessment. In the context of prostate cancer diagnosis and management, DL applications are advancing rapidly, facilitating intelligent analysis and predictive capabilities for clinical tasks such as lesion detection, staging, and prognosis evaluation, thereby supporting clinicians in treatment planning [7–10]. EPE is recognized as an independent predictor of poor prognosis in patients with PCa, strongly linked to adverse outcomes such as positive surgical margins, biochemical recurrence, and distant metastasis following RP. A precise preoperative assessment of EPE is, therefore, essential for informed therapeutic decision-making.

Previous studies have focused on DL models based on mpMRI for preoperative EPE prediction. For instance, Moroianu et al. developed the DL-EPE NET model, which achieved an AUC of 0.80 for EPE prediction with improved sensitivity over the manual interpretation of radiologists [11]. Similarly, Jethro C et al. constructed the SEPERA DL-EPE prediction model, which not only detected EPE but also identified its laterality, demonstrating robust predictive performance (AUC: 0.77) in external, multi-institutional validation, thereby highlighting the stability, safety, and generalizability of the model. Currently, to our knowledge, most of studies assessing EPE deal with single modalities (MRI or PSMA-PET/CT) alone or comparison between them, only few studies have aimed to conduct and integrate mpMRI and PSMA-PET/CT multimodality for EPE evaluation. This is the first study to construct deep learning combined model for EPE prediction, and to find that the combined model achieved superior predictive performance validated with external data [12–15].

In current clinical practice, preoperative determination of EPE primarily depends on the manual interpretation of mpMRI images by radiologists. This process is typically guided by a set of MRI indicators detailed in the PI-RADS reporting system, which includes tumor capsular contact, capsular irregularity and/or bulging, loss of rectoprostatic angle, asymmetry in the neurovascular bundle, invasion of extraprostatic fat or nearby structures (e.g., bladder, rectum, pelvic walls), and seminal vesicle involvement. Research indicates that the accuracy of EPE determination by radiologists using these signs varies, with AUC values ranging from 0.60 to 0.75 [16].

The EPE-grade classification system proposed by Mehralivand offers a simpler alternative to PI-RADS for MRI-based EPE assessment, designed to facilitate both teaching and practical reporting. A study by Park et al. indicated an improved radiologist EPE assessment, with AUC values between 0.77 and 0.85 when utilizing the EPE-grade scoring system. However, in our study, the AUC for EPE assessment with the EPE-grade scoring system was 0.64 at Center 1 and 0.63 at Center 2, which is lower than the results reported in previous studies. This discrepancy may be attributed to factors such as the relatively small sample size in this study, a higher prevalence of cases with microscopic EPE lacking MRI-visible manifestations, and a lower proportion of cases with clear EPE features like seminal vesicle invasion or invasion of adjacent structures.

It is also noteworthy that the EPE-grade scoring system was initially proposed based on a single-center retrospective study, and its predictive performance remains under-validated in multi-center settings. Our findings suggest that the stability and generalizability of the system may require further improvement for robust external validation. Accurately predicting EPE preoperatively remains challenging for radiologists, as distinguishing subtle imaging abnormalities on conventional morphological imaging is often difficult; many signs of EPE are only detectable at the microscopic level.

In recent years, PET/CT, especially PSMA-PET/CT, has gained clinical traction as an innovative imaging technique for patients with PCa, demonstrating considerable use in PCa diagnosis, staging, therapeutic response evaluation, and prognosis assessment [17–19]. A prospective cohort study conducted by Cysouw et al. demonstrated that the 18 F-DCFPyL PET radiomics model exhibited excellent predictive capabilities for EPE in PCa patients [20]. Another study compared the diagnostic efficacy of conventional mpMRI and 68Ga-PSMA PET for EPE evaluation in PCa. The findings revealed that mpMRI exhibited limited sensitivity but high specificity compared to PSMA PET/CT in assessing EPE [21]. Recently, there is a growing interest in the field of deep learning for medical image analysis. DL algorithms can automatically extract high-throughput quantitative parameters that visual assessments may overlook. This allows for the extraction of more tumor-related information and achieves better predictive performance. Based on these advancements, it was hypothesized that using deep learning (DL) algorithms for high-throughput image data extraction and utilizing combined information from mpMRI and PSMA-PET/CT multimodal imaging to develop predictive models could help radiologists in detecting microscopic EPE, which remains challenging to discern on conventional morphological imaging. This study, therefore, constructed a radiologist EPE-grade scoring system enhanced by a multimodal imaging DL model.

The findings indicate that the DL model significantly elevated the diagnostic accuracy of radiologists in predicting EPE, with AUC values rising from 0.64 to 0.82 in the training and internal testing groups of Center 1. This improvement was validated in the external validation group (Center 2), where the AUC increased from 0.63 to 0.87, corroborating the initial hypothesis and enhancing the accuracy of preoperative EPE assessments by radiologists. Hou et al. conducted a dual-center retrospective study in China to predict EPE using a DL model (PAGNet model) constructed from mpMRI images of 849 patients with PCa undergoing radical prostatectomy [22]. DL-assisted interpretation revealed superior diagnostic performance compared to radiologist interpretation alone, achieving AUC values of 0.86 and 0.81 in the training and internal validation sets, respectively, though performance decreased to 0.73 in the external validation set. These results propose that DL-based image mining and model development may complement and refine existing radiologist assessment protocols, enhancing accuracy in EPE evaluation and providing a promising AI-assisted precision diagnostic approach for clinical practice. On the other hand, it should also be noted that there remains a small number of cases where the AI model fails to make accurate prediction. The possible reasons may lie in the variability of observer expertise and inconsistencies in image quality.

This study has several limitations. As a retrospective analysis, despite the inclusion of an external independent validation cohort, the dataset size remains limited. This limitation warrants further assessment through larger, prospective multicenter studies to better assess model accuracy, stability, and generalizability. Additionally, the sample sizes for both model construction and external validation remain modest, indicating a need for further expansion to validate the findings. Moreover, the interpretability of the DL model is limited; future research could address this limitation by employing techniques such as heat maps to visualize the decision-making process of the model [23]. At last, the study population had relatively high PSA levels which may be little far from real-life practice.

Conclusion

In conclusion, this study demonstrated that a deep residual convolutional neural network (ResNet-50) effectively predicts EPE by leveraging multimodal imaging from mpMRI and PSMA-PET/CT. The integration of the multimodal imaging DL model- notably enhanced the accuracy of radiologists in assessing EPE, exhibited the potential to serve as a reliable and robust tool for clinicians in making more individualized and precise therapeutic decisions for patients with PCa.

Abbreviations

Pca

Primary prostate cancer

EPE

Extraprostatic extension

PSA

Prostate-specific antigen

GS

Gleason score

SVI

Seminal vesicle invasion

BMI

Body mass index

MRI

Magnetic resonance imaging

PET

Positron emission tomography

PSMA

Prostate-specific membrane antigen

RF

Radiomics feature

IBSI

Image biomarker standardization initiative

ROI

Region of interest

CI

Confidence interval

NRI

Net reclassification improvement

ROC

Receiver operating characteristic curve

AUC

Area under curve

ACC

Accuracy

SEN

Sensitivity

SPE

Specificity

PPV

Positive predictive value

NPV

Negative predictive value

DCA

Decision curve analysis

SUVmax

Maximum standardized uptake value

Author contributions

Fei Yao: Conceptualization, Software, Writing – original draft. Heng Lin: Data curation, Software. Ying-Nan Xue: Data curation, Formal Analysis. Yuan-Di Zhuang: Data curation, Formal Analysis. Shu-Ying Bian: Data curation, Formal Analysis. Ya-Yun Zhang: Formal Analysis, Visualization. Yun-Jun Yang: Conceptualization, Funding acquisition. Ke-Hua Pan: Conceptualization, Writing – original draft, Writing – review & editing. All authors read and approved the final draft.

Funding

This study was funded by the Key Laboratory of Novel Nuclide Technologies on Precision Diagnosis and Treatment & Clinical Transformation of Wenzhou City (Grant Number: 2023HZSY0012). The funding body had no role in the design of the study and collection, analysis, and interpretation of data and in writing the manuscript.

Data availability

The datasets used and/or analyzed during the current study are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

This study was conducted with approval from the Ethics Committee of The First Affiliated Hospital of Wenzhou Medical University (Approval Number: KY2024-R181). This study was conducted in accordance with the declaration of Helsinki. Informed consent was waived due to the retrospective nature of the study, which involved the analysis of existing clinical data with no direct patient contact or intervention.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.