Deep learning-based diffusion-weighted imaging vs. conventionally obtained diffusion-weighted imaging in prostate cancer extracapsular extension detection: a multicenter retrospective study

Abstract

Background

Extracapsular extension (ECE) of prostate cancer (PCa) is a crucial determinant for preoperative staging, significantly influencing urologists’ surgical planning. While magnetic resonance imaging (MRI) aids staging by depicting ECE, its diagnostic accuracy is frequently constrained by poor image quality. Deep learning-based diffusion-weighted imaging (DL-DWI) holds potential for enhancing diagnostic accuracy through improved image quality.

Methods

This multicenter retrospective study enrolled 252 consecutive patients with clinically suspected PCa from five centers between June 2019 and September 2023. All patients underwent multiparametric MRI (mpMRI) followed by radical prostatectomy. A denoising diffusion probabilistic model (DDPM) was developed to synthesize high-b-value DL-DWI images. Two radiologic residents independently evaluated the image quality of DL-DWI and conventionally obtained DWI (C-DWI) using a five-point Likert scale. Objective image quality metrics including estimated signal-to-noise ratio (eSNR), contrast-to-noise ratio (CNR), edge rise distance (ERD), and edge rise slope (ERS) were quantitatively assessed and compared between sequences. Four radiologists (experience range: 2–20 years) independently made the diagnosis of each patient based on the DL-DWI and C-DWI using the four-scale ECE grade system. Intra-reader agreement was assessed using weighted kappa, and inter-reader agreement using Fleiss’ kappa. Diagnostic performance was compared using multi-reader multi-case receiver operating characteristic analysis, with the area under the curve (AUC) as the metric. Decision curve analysis (DCA) evaluated the net clinical benefit. Statistical significance was defined as a two-sided P < 0.05.

Results

DL-DWI demonstrated significantly superior image quality compared to C-DWI (P < 0.05) and higher agreement for both intra-reader (weighted κ: 0.59, 0.45 vs. 0.40, 0.33 for readers 1 and 2, respectively) and inter-reader (weighted κ: 0.59 vs. 0.48) quality assessments. Quantitative analysis confirmed substantial improvement in DL-DWI across all objective metrics: eSNR (56.66 ± 7.36 vs. 31.76 ± 12.97, P < 0.001), CNR (28.89 ± 9.27 vs. 10.53 ± 5.48, P < 0.001), ERD (2.85 [IQR:1.90–3.80] vs. 4.27 [IQR:3.27–5.60], P < 0.001), and ERS (69.48 ± 19.15 vs. 23.82 ± 10.29, P < 0.001). For ECE detection, DL-DWI showed higher intra-reader agreement (weighted κ for readers 3–6: 0.53, 0.56, 0.67, 0.70 vs. 0.53, 0.34, 0.52, 0.49) and inter-reader agreement (Fleiss κ: 0.25 vs. 0.18) compared to C-DWI. DL-DWI achieved significantly higher diagnostic performance (mean AUC: 0.76 [95% CI: 0.71–0.80] vs. 0.70 [95% CI: 0.65–0.74], P < 0.001). DCA indicated greater net clinical benefit for DL-DWI compared to C-DWI.

Conclusion

DL-DWI appears to outperform C-DWI in both image quality and clinical utility, showing potential to improve ECE diagnostic accuracy for residents, enhance inter-reader consistency, and provide greater net benefit in clinical decision-making.

Clinical trial number

Not applicable.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12880-025-02109-x.

Introduction

Prostate cancer (PCa) is the second most common cancer in men and the third most common cancer of all cancer types [1]. Accurate preoperative staging of PCa is critical for guiding treatment decisions [2]. Extracapsular extension (ECE) is a key determinant in preoperative staging and significantly influences urologists’ surgical planning [3]. Current guidelines recommend neurovascular bundle (NVB) resection in patients with ECE to reduce positive surgical margins; however, this procedure can substantially compromise urinary continence and sexual function [2, 4]. Therefore, accurate preoperative detection of ECE is essential for urologists to tailor surgical management, particularly in determining the feasibility of nerve-sparing techniques [5]. However, there are still several challenges in the current diagnosis of ECE.

Magnetic resonance imaging (MRI) is a well-established modality for localizing suspected PCa lesions, primarily to guide targeted biopsies. Beyond tumor detection, MRI also plays a role in preoperative local staging, including the assessment of ECE [6]. Previous studies have reported that MRI outperforms clinical nomogram in predicting ECE, and the combined use of MRI and clinical nomogram further improves the prediction of ECE [7–9]. However, the diagnostic performance of MRI for ECE exhibits limitations, with reported sensitivities ranging heterogeneously from 30% to 70% [7, 10, 11]. This heterogeneity can be attributed to multiple factors, including variations in MRI equipment, image quality, sequence selection, radiologist experience, and standardized reporting criteria [12, 13].

Diffusion-weighted imaging (DWI) is a key component of multiparametric MRI (mpMRI) in the detection and characterization of PCa. As the b-value increases, DWI enables detection of small changes in diffusion because of better contrast between malignant and background tissue [6, 14–16]. However, high b-value (>1000 s/mm²) DWI is technically challenging to obtain in clinical practice; conventionally obtained DWI (C-DWI) frequently suffers from poor in-plane spatial resolution, distortion, and artifacts. Moreover, an increase in the b-value is generally accompanied by a significant reduction in the signal-to-noise ratio and requires a prolonged acquisition time [17, 18]. With the development of deep learning technology, deep learning-based DWI (DL-DWI) demonstrates fewer distortion artifacts, and better spatial resolution [16, 19–21]. Recent studies have shown that DL-DWI significantly improves diagnostic accuracy and inter-radiologist consistency, but these studies have primarily focused on the preoperative assessment of PCa malignancy [22, 23]. To our knowledge, the potential benefit of DL-DWI for local staging in patients with PCa is still unclear.

Therefore, this multicenter retrospective study aims to compare the diagnostic performance of C-DWI and DL-DWI for the preoperative detection of ECE in patients with PCa.

Methods

This retrospective multicenter study has been reported in line with the Standards for Reporting Diagnostic Accuracy Studies (STARD) [24]. It was approved by the Medical Ethics Committee of Guangdong Provincial People’s Hospital (approval No. KY2023-146-01), which waived the requirement for informed consent due to the retrospective and anonymous nature of the data analysis. All procedures were conducted in accordance with the ethical standards of the 1964 Helsinki Declaration and its later amendments.

Patients

We reviewed all consecutive patients undergoing pelvic phased-array prostatic mpMRI examination on five 3.0 T MRI scanners for primary staging of PCa between June 2019 and September 2023 from five centers. The inclusion criteria were as follows: (1) Patients who underwent radical prostatectomy and had available the requisite mpMRI sequences according to Prostate Imaging-Reporting and Data System version 2.1 (PI-RADS v2.1), including T2-weighted imaging (T2WI), DWI, and the corresponding apparent diffusion coefficient (ADC) map, acquired within two weeks prior to surgery. (2) Complete clinical information and pathological examination results. Exclusion criteria: Patients who received treatment before the MRI examination (including hormone therapy, radiation therapy, cryotherapy, or surgery).

A total of 313 consecutive patients who underwent radical prostatectomy were initially identified. After excluding 61 patients based on the exclusion criteria, 252 patients were included in the final analysis are shown in Fig. 1. The clinical and pathological characteristics of patients, including age, prostate-specific antigen (PSA) testing, surgical Gleason score (GS), D’Amico risk classification, and tumor stage are summarized in Table 1.

Fig. 1.

Diagram for inclusion of patients into the study

Table 1.

Summary of characteristics in patients

Characteristic	Value
Age (y)a	69 ± 7
PSA level (ng/ml)b	18.34 [9.90-47.09]
Surgical Gleason Score
3 + 3	24 (9.5)
3 + 4	60 (23.8)
4 + 3	59 (23.4)
8	55 (21.9)
≥ 9	54 (21.4)
D’Amico risk group
Intermediate risk	21 (8.3)
High risk	231 (91.7)
Tumor stage
pT2a	23 (9.1)
pT2b	10 (4)
pT2c	104 (41.3)
pT3a	50 (19.8)
pT3b	65 (25.8)
Pathological ECE
Present	115 (45.6)
Absent	137 (54.4)
Pathological SVI
Present	65 (25.8)
Absent	187 (74.2)

Unless otherwise specified, data are number of patients (percentages)

PSA, prostate-specific antigen; ECE, extracapsular extension; SVI, seminal vesicle invasion

^aData are means ± standard deviations

^bData are medians [interquartile ranges]

MRI experiments

All images were acquired using five 3-T MRI scanners from five clinical centers. Signal reception utilized a phased-array 18-channel body coil combined with an integrated 32-channel spine coil. The imaging protocol comprised axial T2WI, axial DWI, with ADC maps derived from both the C-DWI and DL-DWI sets using a mono-exponential model. All high-b-value C-DWI images (b = 1300–1500 s/mm²) used in this study were directly acquired from the scanners, not computationally generated, with the variation in maximum b-values reflecting differences in clinical protocols across centers. The multicenter MRI image acquisition parameters are listed in Table 2.

Table 2.

The magnetic resonance image acquisition parameters of the multicenter

Parameter	Center 1		Center 2		Center 3		Center 4		Center 5
Manufacturer	Philips		General Electric		Siemens		General Electric		Siemens
Model	Ingenia		Signa		Skyra		750 plus		Verio
Field strength (T)	3		3		3		3		3
Sequence name	T2WI	DWI	T2WI	DWI	T2WI	DWI	T2WI	DWI	T2WI	DWI
Field of view (mm²)	200 × 200	112 × 112	250 × 250	210 × 210	200 × 200	281 × 379	200 × 200	240 × 240	90 × 90	84 × 84
Acquisition matrix	320 × 320	103 × 103	240 × 220	96 × 96	320 × 256	178 × 132	200 × 200	128 × 64	320 × 320	182 × 182
Thickness (mm)	3	4.5	3	3	3.5	3	3	3	3	3
Distance fact (mm)	5	5	5	5	3.5	3	3	3	3	3.6
Number of averages	1	1,1,3	1	1,2,4	2	1,3,6	2	1,1,2,4	2	1,2,4
B-value (s/mm²)	NA	50,800, 1300	NA	50,1000, 1500	NA	50,1000, 1500	NA	0,800,1000, 1500	NA	50,1000, 1500
Echo time(msec)	100	62	74	58	101	79	104	64	85	64
Repetition time (msec)	3490	6000	3420	6000	6910	5600	4000	6000	7580	3920
Pixel bandwidth (Hertz perpixel)	138	2152	195	1953	200	2160	325	1953	200	1115
Acquisition time (min: sec)	2:40	2:41	3:01	3:02	2:08	3:05	3:04	3:52	3:12	3:14

T2WI, T2-weighted imaging; DWI, diffusion-weighted imaging

DL-DWI generation

We employed an image-to-image translation methodology based on the denoising diffusion probabilistic model (DDPM) [25] for synthesizing high-b-value DWI. All DICOM datasets were systematically categorized and filtered according to b-values extracted from DICOM tags. Specifically, DWI acquisitions with b-values of 0–50 s/mm² were designated as standard-b-value inputs for the generative model, while those acquired at 1500 s/mm² served as the reference standard for model output.

The preprocessing pipeline incorporated several critical steps. Each standard-b-value and high-b-value image pair initially underwent rigid registration to establish precise spatial and anatomical correspondence of prostate volumes. Subsequently, all registered images were systematically cropped to 112 × 112 pixels based on prostate gland dimensions. Finally, intensity resampling was performed to normalize all voxel values to the standardized range of [-1, 1].

For the network architecture, we deployed a DDPM-based image synthesis framework built upon a U-Net backbone. The model was optimized to execute iterative denoising across multiple noise scales, initiating from Gaussian noise distributions. To overcome the inherent stochasticity in conventional DDPM generation, we implemented a conditional diffusion paradigm wherein standard-b-value DWI served as guiding information, enabling controlled synthesis of high-b-value DWI from random noise inputs.

The diffusion process was parameterized with 2000 timesteps, with the noise schedule (where ) linearly annealed from to throughout both diffusion and reverse processes. Model optimization employed the Adam algorithm with a learning rate of and batch size of 32. The complete implementation was developed in Python 3.11 using PyTorch 2.6, executed on an Ubuntu 22.04 platform with NVIDIA RTX 4090 GPU acceleration.

Image quality analysis

Two radiologic residents (reader 1: T.S. and reader 2: L.Z.) evaluated all C-DWI and DL-DWI images in random order. To evaluate the image quality (in-folding, ghosting, susceptibility, blurring, etc.), the image quality of both DWI sequences was rated visually on a five-point Likert scale: 1: very poor image quality, non-diagnostic; 2: poor image quality, significantly impacting diagnosis; 3: moderate image quality, mildly impacting diagnosis; 4: good image quality, not impacting diagnosis; 5: excellent image quality.

After 2 weeks, the same image quality analysis was performed again by the two readers in a different order.

To quantitatively evaluate the image quality of both C-DWI and DL-DWI sequences, objective metrics were systematically measured. Radiologist (J.G.) was asked to draw ROIs on the highest b-value DWI images. Three ROIs were placed for each case: one ROI was positioned in the center of the lesion on the slice with the largest lesion extent; a second reference ROI was placed in the corresponding contralateral normal tissue (or in healthy appearing tissue of the same anatomical zone if the contralateral tissue was abnormal); a third ROI was placed in the center of the bladder to calculate the standard deviation of the noise (SD_noise) in a noise-only area. The mean signal intensities in the lesion ROI (S_lesion), the reference ROI (S_normal), and SD_noise were recorded for subsequent analysis.

The estimated signal-to-noise ratio (eSNR) was calculated as:

The contrast-to-noise ratio (CNR) was calculated as:

Image sharpness was quantitatively assessed by measuring the edge rise distance (ERD) and edge rise slope (ERS) at well-defined tissue boundaries. The ERD was defined as the spatial distance required for the signal intensity to rise from 10% to 90% of the total edge transition. The ERS was subsequently calculated as the ratio of the signal intensity change to the ERD.

Diagnostic assessment

Two 2nd-year radiologic residents (reader 1: T.S. and reader 2: L.Z.) and two radiologists (reader 3: M.D. and reader 4: F.Y.) with more than 15 years of experience in prostate MRI diagnosis blinded to the pathological results and all clinical information performed ECE diagnosis according to a four-scale ECE grading system [7]. This criteria-based ECE grading system diagnosis ECE based on the capsular contact length (CCL), capsular irregularity and bulge, and frank breach of the capsule. The grading is as follows: grade 0: no suspicion for pathological ECE; grade 1: either CCL of 15 mm or greater or capsular irregularity or bulge; grade 2: both CCL of 15 mm or greater and capsular irregularity or bulge; grade 3: frank ECE visible at mpMRI. DWI served as the primary standard, with T2WI and ADC map used as auxiliary references. The cases of the ECE grading system are shown in Fig. 2. For analysis, grades 2 or 3 were considered a positive test result for ECE, while grades less than 2 were considered a negative test result.

Fig. 2.

Representative axial multiparametric MRI (mpMRI) sequences demonstrate the prostate extracapsular extension (ECE) grading system. Images are arranged from left to right as: T2-weighted imaging, diffusion-weighted imaging, and apparent diffusion coefficient maps. Grade 0: no suspicion for pathological ECE (arrows); Grade 1: capsular contact length ≥ 15 mm and/or capsular irregularity/bulge (arrows); Grade 2: both capsular contact length ≥ 15 mm and capsular irregularity/bulge (arrows); Grade 3: definite extracapsular extension visible at mpMRI (arrows)

All images were anonymized, shuffled, and randomly assigned into two groups. In each group, only one of the two DWI sets for each patient was included. To minimize recognition bias, the same prostate ECE evaluation was repeated by the four readers in a different order after 2 weeks.

Statistical analysis

The normality of continuous variables was assessed using the Kolmogorov-Smirnov test. Continuous variables with a normal distribution are presented as mean ± standard deviation (SD), while those with a non-normal distribution are presented as median and interquartile range (IQR). Categorical variables are reported as frequencies and percentages. Group comparisons for continuous variables were performed using the t-test or Mann-Whitney U test, depending on the distribution. The Chi-square test and McNemar’s test were used for comparing categorical variables. Intra-reader agreement was assessed using weighted kappa, while inter-reader agreement was evaluated using Weighted kappa (for two readers) and Fleiss kappa (for more than two readers). The κ coefficients were assessed as follows: 0.01–0.20, slight agreement; 0.21–0.40, fair agreement; 0.41–0.60, moderate agreement; 0.61–0.80, substantial agreement; and 0.81–0.99, almost perfect agreement.

A multi-reader multi-case (MRMC) approach including U-statistics was used to calculate the each reader’s area under the receiver operating characteristics (ROC) curve (AUC), sensitivity, and specificity, all reported with 95% confidence intervals (95% CI) [26]. Delong test was applied to statistically compare the AUC values between C-DWI and DL-DWI. The clinical usefulness and net benefits of the models were assessed using a decision curve analysis (DCA). All statistical analysis was performed in R (version 4.3.1; https://www.R-project.org/). The reported statistical significance levels were all two-sided, with statistical significance set at P value < 0.05.

Results

Mean patient age was 69 ± 7 years, median PSA value was 18.34 ng/ml [IQR: 9.90-47.09]. 45.6% (115/252) patients were diagnosed with ECE and the remaining 54.4% (137/252) were non-ECE. 91.7% (231/252) of the patients had a high-risk setting, and 8.1% (21/252) had an intermediate risk setting according to the D’Amico classification. The ECE group exhibited higher PSA level than the non-ECE group (36.07 ng/ml [IQR: 15.74–65.40] vs. 12.24 ng/ml [IQR: 6.42–24.99], P < 0.001). The distribution of GS differed significantly between the ECE and non-ECE groups, and the ECE group had less GS6 and more GS6 + than the non-ECE group (P < 0.001). There was no statistically significant difference in the age of patients between the ECE group and non-ECE group (69 ± 7 years vs. 70 ± 7 years; P = 0.35).

Intra- and Inter-reader agreement in DWI sets for image quality analysis and ECE detection

Compared with C-DWI, DL-DWI yielded higher intra-reader agreement for the image quality analysis and two readers got moderate intra-reader agreement (weighted κ reader 1: 0.59 vs. 0.40 and reader 2: 0.45 vs. 0.33). This indicates that DL-DWI has better intra-reader consistency than C-DWI. In addition, DL-DWI also showed higher inter-reader agreement than C-DWI for the two readers in the image quality analysis (weighted κ: 0.59 vs. 0.48).

The intra-reader agreement of ECE evaluation using different DWI sets are shown in Table 3. By using the C-DWI, three readers got moderate intra-reader agreements (weighted κ: 0.53, 0.52, and 0.49, respectively) and one reader got fair intra-reader agreements (weighted κ: 0.34), while two readers got moderate intra-reader agreement (weighted κ: 0.53, 0.56, respectively), and two readers got substantial intra-reader agreement (weighted κ: 0.67, 0.70, respectively) with the use of DL-DWI.

Table 3.

The intra-reader agreement of ECE evaluation using different DWI sets

Reader	Grade 0	Grade 1	Grade 2	Grade 3	Weighted kappa
Reader 1
C-DWI  DL-DWI	48(19.0) 74(29.4)	118(46.8) 86(34.1)	62(24.6) 51(20.2)	24(9.6) 41(16.3)	0.53 0.53
Reader 2
C-DWI  DL-DWI	9(3.6) 41(16.3)	141(56.0) 110(43.7)	60(23.8) 61(24.2)	42(16.6) 40(15.8)	0.34 0.56
Reader 3
C-DWI  DL-DWI	45(17.9) 22(8.7)	64(25.4) 78(31.0)	83(32.9) 77(30.6)	60(23.8) 75(29.7)	0.52 0.67
Reader 4
C-DWI  DL-DWI	22(8.7) 23(9.1)	61(24.2) 66(26.2)	95(37.7) 90(35.7)	74(29.4) 73(29.0)	0.49 0.70

Data in parentheses are percentages. C-DWI, conventional obtained diffusion-weighted imaging; DL-DWI, deep learning-based diffusion-weighted imaging

For inter-reader agreement of ECE evaluation using different DWI sets, DL-DWI demonstrated higher consistency metrics than C-DWI across all reader groupings (Supplementary Table S1). The Fleiss’ κ for all four readers was 0.25 for DL-DWI compared to 0.18 for C-DWI. When analyzed by experience level, the weighted κ between the two radiologic residents increased from 0.25 with C-DWI to 0.41 with DL-DWI. Similarly, the agreement between the two senior radiologists improved from κ = 0.33 to 0.37. The agreement between the combined resident group and senior group showed an increase from κ = 0.32 to 0.40.

Image quality analysis

The evaluation of image quality showed that there were significant differences in scores between the C-DWI and DL-DWI (Reader 1: χ2 = 23.8; Reader 2: χ2 = 13.6; both P < 0.05); compared with DL-DWI, C-DWI yielded more scores of 1 to 3 and fewer scores of 4 and 5, indicating poor image quality on C-DWI, which may impact the diagnosis to a certain extent. The mean scores for DL-DWI were also significantly higher than those for C-DWI (Reader 1: 4.6 ± 0.2 vs. 3.2 ± 0.7; Reader 2: 4.6 ± 0.3 vs. 3.1 ± 0.5; both P < 0.05), which indicates that DL-DWI provides better image quality than C-DWI. These subjective evaluation results are summarized in Supplementary Table S2.

Consistent with the subjective evaluation, objective quantitative metrics further confirmed the superiority of DL-DWI. The eSNR was significantly higher in DL-DWI compared to C-DWI (56.66 ± 7.36 vs. 31.76 ± 12.97, P < 0.001). Similarly, the CNR demonstrated marked improvement in DL-DWI (28.89 ± 9.27 vs. 10.53 ± 5.48, P < 0.001). Quantitative assessment of image sharpness revealed that DL-DWI had significantly shorter ERD (2.85 [IQR:1.90–3.80] vs. 4.27 [IQR:3.27–5.60], P < 0.001) and steeper ERS (69.48 ± 19.15 vs. 23.82 ± 10.29, P < 0.001) compared to C-DWI, indicating superior spatial resolution and edge sharpness in the DL-DWI images. These objective measurements are detailed in Supplementary Table S2.

Diagnostic accuracy evaluation for ECE detection

Examples of patients with a positive ECE and non-ECE in different DWI sets are shown in Fig. 3, among the two DWI sets, the DL-DWI demonstrated the clearer contours of the prostate and lesions, the higher suppression of normal prostate tissue, and the less noise and the fewer artifacts as assessed by using visual perception.

Fig. 3.

T2-weighted images (T2WI), apparent diffusion coefficient (ADC) maps, and calculated ADC (CAL-ADC) maps were derived from deep learning-based diffusion-weighted imaging (DL-DWI) using a mono-exponential model, images from the conventionally obtained diffusion-weighted imaging (C-DWI) show high signal intensity in the tumor (arrows), images from the DL-DWI generated by the deep learning model based on the diffusion model show higher image quality and clarity of the tumor (arrows), images from the fusion were created by combining DWI and T2WI. Patient A: A case with pathologically confirmed negative extracapsular extension of the prostate and a prostate-specific antigen level of 6.73 ng/ml, and DWI b-value = 1000 s/mm² (center 2), two of four readers rated the C-DWI as negative for extracapsular extension, whereas all readers rated the DL-DWI as negative for extracapsular extension. Patient B: A case with pathologically confirmed negative extracapsular extension of the prostate and a prostate-specific antigen level of 7.71 ng/ml, and DWI b-value = 1500 s/mm² (center 4), two radiologic resident readers rated the C-DWI as negative for extracapsular extension, whereas all readers rated the DL-DWI as negative for extracapsular extension. Patient C: A case with pathologically confirmed positive extracapsular extension of the prostate and a prostate-specific antigen level of 10.20 ng/ml, and DWI b-value = 1300 s/mm² (center 1), one of four readers rated the C-DWI as positive for extracapsular extension, whereas three of four readers rated the DL-DWI as positive for extracapsular extension. Patient D: A case with pathologically confirmed positive extracapsular extension of the prostate (seminal vesicle invasion) and a prostate-specific antigen level of 107 ng/ml, and DWI b-value = 1500 s/mm² (center 3), all of four readers rated the C-DWI and DL-DWI as positive for extracapsular extension

The comparison of diagnostic accuracy of different DWI sets based on the ECE grading system is summarized in Fig. 4; Table 4, the mean AUC for ECE detection was significantly higher with DL-DWI than with C-DWI (mean AUC: 0.76 [95% CI: 0.71–0.80] vs. 0.70 [95% CI: 0.65–0.74], P < 0.001). Compared to C-DWI, the AUC of DL-DWI for the two radiologic residents had a significant improvement (P < 0.05), for the two senior radiologists, there were no differences in the AUCs among the two DWI sets.

Fig. 4.

Multi-reader multi-case receiver operating characteristic curve comparison of diagnostic accuracy of conventionally obtained diffusion-weighted imaging and deep learning-based diffusion-weighted imaging based on prostate extracapsular extension grading system

Table 4.

Reader diagnostic performance of C-DWI and DL-DWI for the detection of ECE

Parameter and Reader	C-DWI	DL-DWI	P Value
AUC
Reader1  Reader2  Reader3  Reader4  Mean	0.63 (0.57–0.70) 0.71 (0.65–0.77) 0.72 (0.66–0.78) 0.73 (0.67–0.79) 0.70 (0.65–0.74)	0.74 (0.68–0.80) 0.78 (0.73–0.83) 0.73 (0.67–0.79) 0.76 (0.70–0.82) 0.75 (0.71–0.80)	0.005 0.004 0.77 0.34 < 0.001
Sensitivity (%)
Reader1  Reader2  Reader3  Reader4  Mean	45 (36, 54) [52/115] 60 (51, 69) [69/115] 70 (61, 78) [80/115] 82 (75, 89) [94/115] 64 (58, 70) [74/115]	58 (49, 67) [67/115] 66 (57, 75) [76/115] 65 (57, 74) [75/115] 84 (77, 90) [96/115] 68 (62, 73) [78/115]	0.02 0.12 0.30 0.60 0.04
Specificity (%)
Reader1  Reader2  Reader3  Reader4  Mean	75 (68, 82) [103/137] 76 (69, 83) [104/137] 62 (54, 70) [85/137] 45 (37, 54) [62/137] 65 (59, 70) [89/137]	82 (75, 88) [112/137] 82 (75, 82) [112/137] 71 (63, 78) [97/137] 51 (43, 60) [70/137] 71 (67, 76) [98/137]	0.15 0.14 0.04 0.21 0.002

Data in parentheses are 95% confidence intervals; data in brackets are raw data. The P values for comparing AUCs between C-DWI and DL-DWI were calculated using the DeLong test, while McNemar’s test was used for comparing sensitivity and specificity. C-DWI, conventional obtained diffusion-weighted imaging; DL-DWI, deep learning-based diffusion-weighted imaging; ECE, extracapsular extension; AUC, the area under the receiver operating characteristics curve

Clinic net benefits evaluation for ECE detection

As shown in Fig. 5, across most threshold probability ranges, Diagnoses based on DL-DWI demonstrates a significantly higher net benefit than those based on C-DWI, especially in the medium to high threshold probability ranges. This suggests that the predictive performance of the C-DWI model is not as effective as the DL-DWI model at these thresholds.

Fig. 5.

Decision curve analysis (DCA) for the four readers using conventional (C-DWI) and deep learning-based (DL-DWI) diffusion-weighted imaging. The net benefit is plotted across a range of clinically relevant threshold probabilities. (a) Reader 1, (b) Reader 2, (c) Reader 3, and (d) Reader 4. The DCA evaluates the clinical utility of each method by quantifying the standardized net benefit, which incorporates the trade-offs between true positives and false positives. For example, at a risk threshold of 50% (representing a clinical scenario where intervening on a suspected extracapsular extension is considered justified when the perceived risk is at least 50%), the net benefits for DL-DWI were 0.37, 0.44, 0.30, and 0.39 for Readers 1–4, respectively, consistently higher than the corresponding values of 0.18, 0.31, 0.21, and 0.37 for C-DWI. This indicates that using DL-DWI would provide greater clinical net benefit across this decision threshold. The “All” curve represents the net benefit if all patients are considered positive for extracapsular extension, while the “None” curve represents the net benefit if all patients are considered negative

Discussion

In this study, we compared the diagnostic accuracy of ECE in PCa patients using C-DWI and DL-DWI. DL-DWI improved radiologists’ diagnostic performance in ECE detection, potentially aiding urologists in refining preoperative staging and surgical planning.

Our study results indicate that DL-DWI outperforms C-DWI in the ECE detection capability of radiologists. The mean AUCs of DL-DWI and C-DWI are respectively 0.76 (95% CI: 0.71–0.80) and 0.70 (95% CI: 0.65–0.74) with P < 0.001. By comparing the AUC of two DWI sets from each individual reader, it was found that the DL-DWI did improve the ECE detection capability of radiologic residents, however, for the two senior radiologists the AUCs among the two DWI sets were not significant. This indicates that the image quality of DWI sets may have a greater impact on the diagnostic accuracy of radiologists, which is consistent with previous studies [23, 27], particularly in radiologic residents who are relatively inexperienced with diagnosis [28]. This demonstrates that DL-DWI can help increase the performance of ECE detection more in radiologic residents than in senior radiologists. Several factors may contribute to this observed phenomenon. Senior radiologists may have developed more established diagnostic paradigms that enable them to extract critical features effectively even from conventional images with suboptimal quality, potentially creating a performance ceiling effect. In contrast, radiologic residents might depend more heavily on high-quality image visualization as they are still developing their interpretive skills and diagnostic confidence. The noise suppression and improved anatomical clarity offered by DL-DWI could specifically address the technical limitations that tend to more significantly challenge less experienced readers. Additionally, senior radiologists might approach novel imaging techniques with greater interpretive caution, while radiologic residents may be more inclined to utilize the intuitive advantages of deep learning-generated images.

Our study results also indicate that compared to the C-DWI, we observed higher intra- and inter-reader consistency in DL-DWI. Previous studies have also found that C-DWI has large inter-reader variances, which is consistent with our results and may be caused by poor image quality [19, 29], but they neither compared radiologists with different experiences nor suggested a solution strategy. In this study, the DL-DWI we generated has better image quality, reduced distortion, and fewer noises compared to C-DWI, thus enhancing diagnostic confidence and intra- and inter-reader agreement for ECE detection.

DL-DWI also provides a significantly higher net benefit across most threshold probability ranges compared to the C-DWI, with the difference being particularly pronounced in the medium to high threshold probability ranges. In clinical practice, this means that the DL-DWI is better equipped to correctly diagnose patients who are truly at risk of ECE, particularly in scenarios where higher threshold probabilities are used to make treatment decisions. This improved diagnostic accuracy can lead to more targeted and effective interventions, reducing the likelihood of unnecessary treatments and associated side effects for patients who are not actually at risk of ECE. Moreover, the ability of the DL-DWI model to maintain a high net benefit across various threshold probabilities suggests its robustness and reliability in different clinical settings. Overall, these findings underscore the potential of the DL-DWI to enhance clinical decision-making processes, leading to better patient outcomes and more efficient use of medical resources.

Our study has several limitations that should be acknowledged. First, the retrospective design and relatively limited sample size may introduce potential selection bias, requiring validation through larger prospective studies. Second, the multicenter nature led to variations in imaging parameters across scanners, with some protocols deviating from current PI-RADS v2.1 recommendations, which may affect generalizability despite standardized preprocessing. Third, although DL-DWI demonstrated improved diagnostic accuracy, substantial heterogeneity persisted in radiologists’ subjective assessments, underscoring the need for further standardization and validation of its clinical utility.

Conclusion

Overall, DL-DWI appears to outperform C-DWI in both image quality and clinical utility, showing potential to improve ECE diagnostic accuracy for residents, enhance inter-reader consistency, and provide greater net benefit in clinical decision-making.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

ECE

Extracapsular Extension

PCa

Prostate Cancer

MRI

Magnetic Resonance Imaging

DL-DWI

Deep Learning-based Diffusion-Weighted Imaging

C-DWI

Conventionally obtained Diffusion-Weighted Imaging

DDPM

Denoising Diffusion Probabilistic Model

T2WI

T2-Weighted Imaging

ADC

Apparent Diffusion Coefficient

PI-RADS

Prostate Imaging-Reporting and Data System

PSA

Prostate-Specific Antigen

GS

Gleason Score

AUC

Area Under the Curve

ROC

Receiver Operating Characteristic

DCA

Decision Curve Analysis

MRMC

Multi-Reader Multi-Case

IQR

Interquartile Range

eSNR

Estimated Signal-to-Noise Ratio

NVB

Neurovascular Bundle

CNR

Contrast-to-Noise Ratio

ERD

Edge Rise Distance

ERS

Edge Rise Slope

Author contributions

J.G., T.S., L.Z., M.D, J.C, Y.W., H.L, L.S.: Data curation and Formal analysis; J.G., L.H., F.Y., Y.Z.: Conceptualization; J.G., L.H.: Writing – original draft; J.G., L.H., F.Y.: Writing review and editing.

Funding

This work was supported by the National Science Foundation for Young Scientists of China (No.82302130), Hubei Provincial Natural Science Foundation and Xiangyang-of China (No. 2025AFD041), Hubei Provincial Natural Science Foundation of China (No. 2024AFB431), and the School-level projects of Hubei Polytechnic University (No. 24xjz01W).

Data availability

We declared that materials described in the manuscript, including all relevant raw data, will be freely available to any scientist wishing to use them for non-commercial purposes, without breaching participant confidentiality. The datasets used in this study are available from the corresponding author, Lei Hu, upon reasonable request.

Declarations

Ethics approval and consent to participate

I confirm that I have read the Editorial Policy pages. This study was approved by the Medical Ethics Committee of Guangdong Provincial People’s Hospital (approval No. KY2023-146-01). The requirement for informed consent was waived by the ethics committee due to the retrospective and anonymous analysis of the data. All procedures performed in studies involving human participants were in accordance with the ethical standards of the 1964 Helsinki Declaration and its later amendments.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.