Skip to main content

A transfer learning nomogram for predicting prostate cancer and benign conditions on MRI



Deep learning has been used to detect or characterize prostate cancer (PCa) on medical images. The present study was designed to develop an integrated transfer learning nomogram (TLN) for the prediction of PCa and benign conditions (BCs) on magnetic resonance imaging (MRI).


In this retrospective study, a total of 709 patients with pathologically confirmed PCa and BCs from two institutions were included and divided into training (n = 309), internal validation (n = 200), and external validation (n = 200) cohorts. A transfer learning signature (TLS) that was pretrained with the whole slide images of PCa and fine-tuned on prebiopsy MRI images was constructed. A TLN that integrated the TLS, the Prostate Imaging–Reporting and Data System (PI-RADS) score, and the clinical factor was developed by multivariate logistic regression. The performance of the TLS, clinical model (CM), and TLN were evaluated in the validation cohorts using the receiver operating characteristic (ROC) curve, the Delong test, the integrated discrimination improvement (IDI), and decision curve analysis.


TLS, PI-RADS score, and age were selected for TLN construction. The TLN yielded areas under the curve of 0.9757 (95% CI, 0.9613–0.9902), 0.9255 (95% CI, 0.8873–0.9638), and 0.8766 (95% CI, 0.8267–0.9264) in the training, internal validation, and external validation cohorts, respectively, for the discrimination of PCa and BCs. The TLN outperformed the TLS and the CM in both the internal and external validation cohorts. The decision curve showed that the TLN added more net benefit than the CM.


The proposed TLN has the potential to be used as a noninvasive tool for PCa and BCs differentiation.

Peer Review reports


In men, prostate cancer (PCa) is the most frequently diagnosed cancer in 112 of 185 countries globally and is the leading cause of cancer death in 48 countries [1]. The burden of PCa continues to increase in some developing countries [2]. An accurate diagnosis is the first step towards better management for patients with suspicious prostate abnormalities.

In the traditional PCa diagnostic pathway, patients with elevated serum prostate-specific antigen (PSA) and abnormal digital rectal examination (DRE) often undergo transrectal ultrasound-guided biopsy for PCa detection. However, it may cause unnecessary biopsies or missed PCa detection [3, 4]. In recent years, magnetic resonance imaging (MRI) has begun to play an essential role in the detection and diagnosis of PCa [5, 6], and multiparametric magnetic resonance imaging (mp-MRI) before biopsy has been recommended by clinical guidelines [7, 8]. The Prostate Imaging Reporting and Data System (PI-RADS) [9], as a guideline designed to standardize the acquisition, interpretation and reporting of prostate MRI, has been widely used in clinical practice and has shown good performance in the detection of PCa or clinically significant prostate cancer (csPCa) [10, 11]. However, a wide range of benign conditions (BCs) and anatomic patterns show overlapping characteristics with PCa on MRI [12, 13]. Even when utilizing PI-RADS, the assessment accuracy of PCa still varies across radiologists and requires much expertise and experience [14, 15]. In previous studies, the sensitivity and specificity for PCa diagnosis ranged from 73 to 100% and 8–100%, respectively [10]. For patients with PCa or BCs, the management strategies and treatments are different. Therefore, a more accurate and objective approach for PCa diagnosis before biopsy is needed.

Machine learning (ML) and deep learning (DL) applications in radiology have the potential to diminish inter-reader variability, improve radiology workflow and increase radiologist productivity [16]. In previous studies, the sensitivity and specificity for PCa diagnosis using ML ranged from 0.62 to 0.99 and 0.51 to 0.98, respectively [17]. Some studies have shown that the performance of ML and DL models for detecting or characterizing PCa on MRI was comparable to or better than that of some radiologists using PI-RADS [18,19,20]. However, there are some limitations; for example, some models lack external validation, and in ML and DL models, insufficient training examples may cause overfitting. Transfer learning (TL), as one of the strategies to solve overfitting [21, 22], has been adopted by some researchers for PCa detection [23] or classification [19] on MRI and performed better than deep learning or ML models without TL. However, there might be limitations in terms of generalizability due to the relatively small dataset scale or the lack of external validation.

The present study was designed to develop a prediction model incorporating MRI-based TL features, clinical MRI interpretation, and conventional clinical predictors to differentiate PCa from BCs.


Study population

This retrospective study was approved by our institutional review board, and informed consent was waived. A total of 709 consecutive patients with pathologically proven PCa or BCs between 2015 and 2021 were enrolled from two institutions. The inclusion criteria were as follows: (a) pathologically confirmed primary prostate adenocarcinoma or benign conditions; (b) no radical prostatectomy, radiotherapy, hormonal therapy, or other therapies before MRI and transrectal ultrasound-guided biopsy; and (c) prebiopsy MRI in our institutions and an interval between MR examination and biopsy of less than two weeks. The exclusion criteria were as follows: (a) severe MRI artefacts and (b) incomplete MRI images or clinical information. The flow chart of the study population is shown in Fig. 1.

Fig. 1
figure 1

Flow chart of the study population

MRI assessment and pathological evaluation

MRI images were interpreted by two board-certified radiologists (reader 1 and reader 2, as “expert” radiologists and “basic” radiologists, respectively, according to the criteria provided in the European Society of Urogenital Radiology (ESUR) and EAU Section of Urologic Imaging (ESUI) consensus statement [24].). The lesions were scored according to the assessment criteria provided in PI-RADS v2.1 [9]. A consensus was reached by discussion in case of disagreement. For cases with more than one suspicious lesion, the one with the highest PI-RADS assessment category was recorded. The volume of the prostate was calculated using ellipsoid formulation ([maximum anteroposterior diameter] × [maximum transverse diameter] × [maximum longitudinal diameter] × 0.52), and the diameters were measured as suggested in PI-RADS v2.1 [9]; that is, the maximum longitudinal diameter and maximum anteroposterior diameter were measured on midsagittal T2-weighted MRI, and the maximum transverse diameter was measured on axial T2-weighted MRI.

The biopsy cores were acquired by transrectal ultrasound-guided systematic biopsies (10 to 12 cores) and, in some cases, with additional MR-targeted biopsy obtained through cognitive guidance (2 to 5 cores). The biopsy cores or prostatectomy specimens were evaluated by board-certified pathologists, and the pathological diagnosis of prostate adenocarcinoma was made according to the World Health Organization (WHO) classification of tumours of the urinary system and male genital organs [25].

MRI images acquisition and regions of interest (ROI) images acquisition

All patients underwent prostate MRI on 3.0 Tesla MR scanners (Ingenia, Phillips/ TrioTim, Siemens) without an endorectal coil. The main parameters of axial T2 weighted imaging (T2WI) for the training and internal validation set were as follows: the echo time (TE) was 80 ms, the repetition time (TR) was 4000 ms, the spacing between slices was 0 mm, the slice thickness was 3 mm, the field of view (FOV) was 180 mm × 180 mm and the voxel size was 1 × 1.1 × 3. The detailed scanning parameters are shown in Supplemental Table S1 and Table S2.

The ROI refers to the whole prostate, and was delineated by three board-certified radiologists on axial T2-weighted images. To meet the requirements of the TL model for training data, we preprocess the data of the input model to 224 × 224 × 3 images. The detailed process of data preprocessing is shown in Supplement A1.

Building the clinical model (CM)

The CM incorporating the clinical factors and the PI-RADS score was built through the following three steps. First, we analysed inter-reader agreements (reader 1 vs. reader 2) of the PI-RADS score using Cohen’s kappa test. Second, statistical tests of the clinical parameters and the PI-RADS score were conducted using the Mann‒Whitney U test or Pearson Chi-square test. The clinical parameters included age, PSA, and prostate volume. Third, factors with statistically significant differences were selected to develop a clinical model by multivariate logistic regression with a stepwise forward selection of variables, according to Akaike’s information criterion values.

Transfer learning feature extraction based on the transfer learning model

A transfer learning model was proposed to extract the robust features of the prostate from MR images. As shown in Fig. 2, the structure of the transfer learning model was composed of a convolution block and multilayer perceptron (MLP) mixer [26]. The convolution layer is used to obtain the local features of the image. The MLP mixer layer is used for information fusion in the spatial domain and channel domain. In this study, all parameters of the pretrained model were trained by the whole slide images (WSIs) of PCa from the Cancer Genome Atlas [27], which was the source of initial weights for all subsequent models. All layers of the models were unfrozen, and the target model was trained by the dataset of MR data. The detailed training parameters of the transfer learning model are described in Supplement A2.

Fig. 2
figure 2

Transfer learning model based on MLP Mixer

Based on the transfer learning model, 28,320 transfer learning features were extracted from each patient. The architecture of the transfer learning model and the process of feature extraction are shown in Supplement A3.

Building a transfer learning signature (TLS) based on transfer learning feature

The following three steps were performed to build the TLS based on the transfer learning feature. First, the significance of each TL feature in differentiating the BCs and PCa groups was determined using the Mann‒Whitney U test. Second, the least absolute shrinkage and selection operator (LASSO) algorithm based on 10-fold cross-validation was used to select significant TL features and develop TLS. In addition, we also use support vector machine (SVM) and extreme learning machine (ELM) to develop TLS at the same time to verify the performance of the model on internal and external validation datasets. Finally, we computed the TLS score by combining the selected TL features linearly and weighting them based on their coefficients.

Development of a transfer learning nomogram (TLN)

Multifactor logistic regression analysis based on the forward stepwise selection method [28] was used to select risk factors from TLS, age, PSA, volume and PI-RADS and construct the TLN as the final prediction model.

Model validation index

To evaluate the performance of each model, the receiver operating characteristic (ROC) curve, area under the curve (AUC), sensitivity, specificity, accuracy, positive probability value (PPV), negative probability value (NPV), DeLong test, integrated discrimination improvement (IDI) and decision curve analysis (DCA) were calculated.

Statistical analysis

According to reference [29], statistical analysis of the clinical factors and clinical MRI interpretation were performed.


The CM construction

The interrater agreement of prostate volume was excellent, and the ICC value was 0.970. The Interrater agreement of PI-RADS score was medium, and the kappa coefficient was 0.578. The PI-RADS score and prostate volume were unanimously approved by the two readers. Data on clinical factors and PI-RADS scores of all cohorts are presented in Table 1. In the training cohort, the BCs and PCa groups differed significantly in factors including age, prostate volume, PI-RADS score, and PSA (p < 0.01, < 0.001, < 0.001, and < 0.001, respectively).

Table 1 Clinical factors and MRI findings of the patients

The clinical model was built using multivariable logistic regression, where age (odds ratio [OR], 1.051; 95% confidence interval [CI], 1.012–1.091; p value = 0.01), prostate volume (OR, 0.988; 95% CI, 0.980–0.995; p value = 0.001), PSA (OR, 2.595; 95% CI, 1.825–3.692; p value < 0.001) and PI-RADS score (OR,13.699; 95% CI, 6.031–31.118; p value < 0.001) were identified as independent risk factors (Table 2).

Table 2 Independent risk factors associated with prostate cancer in the clinical model by multivariate logistic regression

Transfer learning feature selection and TLS construction

To verify which classifier could achieve better performance, we built three TLSs with LASSO, SVM, and ELM classifiers. The value of each TL feature in differentiating the BCs and PCa groups was determined using the Mann‒Whitney U test. In total, 21,900 TL features significantly differed between the BCs and PCa groups in the training cohort. Among these, 73 TL features with nonzero coefficients were selected for inclusion in the TL score calculation formula using LASSO logistic regression (Fig. 3, Supplement A4). The detailed procedure of constructing the TLS with SVM and ELM is shown in Supplement A4.

Fig. 3
figure 3

Selection by LASSO logistic regression. a The selection of tuning parameter (λ) using a 10-fold cross-validation according to the minimum criteria. At the optimal value of λ, the dotted vertical line was plotted. The optimal value of λ was 0.477, and log(λ)= -0.740. b Coefficient profiles determined by lasso logistic regression analysis of features. At log(λ)= -0.740, the dotted vertical line was drawn, including 73 optimal features with non-zero coefficients. LASSO: least absolute shrinkage and selection operator

TLS assessment and comparison

Comparison of the performance of those models revealed that LASSO displayed significantly higher efficiency of diagnosis in the training (AUC = 0.9700 (95% CI, 0.9549–0.9852)), internal validation (AUC = 0.9023 (95% CI, 0.8596–0.9450)), and external validation cohorts (AUC = 0.8697 (95% CI, 0.0.8202–0.9191)) (Table 3; Fig. 4).

Table 3 Performance of the TLS based on LASSO, SVM, ELM classifiers in the training, internal validation, and external validation cohorts
Fig. 4
figure 4

Receiver operating characteristics curves of each set using different classifier. a training cohort, b internal validation cohort, c external validation cohort. LASSO: transfer learning radiomics nomogram; SVM: Support Vector Machine. ELM: Extreme Learning Machine; TLN: transfer learning nomogram; TLS: transfer learning signature

TLN construction and validation

In the multivariable logistic regression analysis, TLS (OR, 120.780; 95% CI, 34.648–421.028; p value < 0.001), age (OR, 1.088; 95% CI, 1.030–1.150; p value = 0.003), and PI-RADS score (OR, 3.472; 95% CI, 1.198–10.058; p value = 0.022) were independent predictors (Table 4). Incorporating these three independent factors, we constructed a combined model to be presented as a transfer learning nomogram (Fig. 5a). Examples of the clinical use of the nomogram are shown in Fig. 6.

Table 4 The parameters of the TLN for BCs and PCa in patients of the training set
Fig. 5
figure 5

Construction of the TLN and DCA of the various diagnostic models. a The presented TLN that incorporates both TL signature and clinical factors. Calibration curves of the TLS in training (b) and two validation cohorts (c), respectively. d DCA of different diagnostic models. The solid gray and black lines indicate the assumption that all and none of the PCa groups are involved, respectively. The threshold probability was defined as the point at which the expected benefit of the treatment was equal to the benefit of avoiding treatment. The results indicated that the TLN provided a greater net benefit than the clinical model and TLS (range 0.01– 0.99). TLN: transfer learning nomogram; TLS: transfer learning signature; DCA: decision curve analysis

Fig. 6
figure 6

Examples of the nomogram in clinical practice. a A patient with raised prostate-specific antigen and Prostate Imaging Reporting and Data System category 3 lesion in left medial posterior peripheral zone. The total score is 52.5, which corresponds to a prostate cancer risk of less than 0.1. The biopsy result was benign. b A patient with raised prostate-specific antigen and Prostate Imaging Reporting and Data System category 3 lesion in left medial posterior peripheral zone. The total score is 85.1, which corresponds to a prostate cancer risk of greater than 0.9. The biopsy result was Gleason grade group 2 prostate cancer. TLS: transfer learning signature

Using the calibration curve, we confirmed a marked connection between the predicted and actual data in the training cohort (Fig. 5b and c).

The diagnostic performance of the CM, TLS, and TLN are shown in Table 5; the ROC curves of these three models are shown in Fig. 7. In the internal and external validation cohorts, respectively, the transfer learning nomogram achieved the best discrimination (AUC, 0.9255; 95% CI, 0.8873–0.9638 and AUC, 0.8766; 95% CI, 0.8267–0.9264, accuracy of 0.8750 and 0.7700, sensitivity of 0.9280 and 0.7273, and specificity of 0.9293 and 0.8354).

Table 5 Performance of the clinical model, TLS, and TLN in the training, internal validation, and external validation cohorts

The confusion matrix of the TLN is presented in Table 6.

Table 6 Confusion matrix for predicted versus actual categories
Fig. 7
figure 7

Comparison of the performance of the CM, TLS and TLN. a Receiver operating characteristics curves of three models in each set; b Diagnostic performance of the three models. TLN: transfer learning nomogram; TLS: transfer learning signature. CM: clinical model

The Delong test and IDI demonstrated that the TLN had significantly superior performance to both the TLS and CM in all validation cohorts (Delong test, P < 0.001, P = 0.0263; IDI = 0.0611, P < 0.001, IDI = 0.1272, P < 0.001, respectively).

The decision curve showed that the TLN added more net benefit than the CM in differentiating the BCs and PCa groups within the range of threshold probability 0.01 to 0.99 (Fig. 5d).


In the present study, we developed and validated an integrated nomogram incorporating useful TL features, clinical MRI interpretation, and clinical predictors for PCa differentiation. By comparison, the TLN outperformed TLS alone or CM in both the internal and external validation cohorts. The proposed TLN could be a better noninvasive diagnostic tool for differentiating PCa and BCs.

PI-RADS score was identified as an independent risk factor in both CM and TLN. In clinical MRI interpretation, PI-RADS score assignment is evaluated by radiologists based on MRI findings. T2-weighted images are often used to display anatomical information, detect abnormalities and evaluate extraprostatic extension. The apparent diffusion coefficient (ADC) map and high b-value images of diffusion-weighted imaging (DWI) are valuable for lesion detection since prostate cancers often have restricted water molecule diffusion. DCE can be used to detect some small cancers with early enhancement. For the peripheral zone, DWI is the primary determining sequence, and the T1-weighted dynamic contrast material–enhanced (DCE) result can be helpful when the DWI score is intermediate. For the transitional zone, the T2-weighted score is the dominant factor, and the DWI score is also useful in atypical transitional zone nodules. As a standardized risk assessment tool for PCa and csPCa, the PI-RADS score has been incorporated by some prediction models [30, 31] and performed better than those without mp-MRI.

In addition to PI-RADS score assignment, PSA, prostate volume, and age were identified as independent risk factors in our CM and were also significant predictors of PCa or csPCa in previous studies [29, 30]. However, these predictors are not cancer specific; benign conditions, such as benign prostatic hyperplasia, may also result in the enlargement of the prostate and an increase in PSA and are more prevalent in ageing men [32], which may explain why CM performed less strongly in both internal and external validation cohorts than TLS and TLN when differentiating PCa and BCs.

During the TLS construction, we applied TL to mitigate the overfitting in the process of model training using small data and extracting robust features. We chose the WSIs of PCa as the source domain dataset for TL because they are medical images containing tumour histopathological information and are more similar to prostate magnetic resonance images than natural images. Literatures have reported the correlation between prostate MRI characteristics and histological conditions presented in histopathological slides [12, 33]. The more similar the source domain data are to the target domain data, the more a small training dataset can make full use of the transfer of learning [34, 35]. Our TL method can autonomously mine features based on the WSIs of PCa from superficial to deep layers of the image through multilayered networks. These features were more relevant to the task, contained more lesion information, and were essential for differentiating PCa and BCs. The proposed TLS also performed well in the internal and external validation cohorts.

The final prediction model TLN was constructed by combining TLS, age, and PI-RADS score. The performance of the TLN was encouraging, with AUCs of 0.9255 and 0.8766 in the two validation cohorts. Furthermore, we found that the performance of TLN was better than that of CM or TLS alone. This may be because not only imaging features but also clinical factors as well as clinical MRI interpretation were taken into account when constructing the TLN.

The current study has some limitations. First, the sample size used in this research was still limited, and further multicentre validation is needed before the TLN can be used routinely in clinical practice. Second, for most patients, the pathological assessment was based on biopsy cores acquired by transrectal ultrasound-guided biopsies, which may lead to the missed diagnosis of some undetected PCa. Third, our model was restricted to differentiating PCa from BCs, and did not encompass the differentiation between csPCa and clinically insignificant prostate cancer. We are considering developing a reliable model to predict csPCa in future work.


In conclusion, we developed a TLN that could be an important risk assessment tool for the differentiation of PCa and BCs, providing valuable assistance in clinical decision-making.

Availability of data and materials

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



Apparent diffusion coefficient


Area under the curve


Benign conditions


Clinical model


Clinically significant prostate cancer


Decision curve analysis


Dynamic contrast material–enhanced


Deep learning


Digital rectal examination


Diffusion-weighted imaging


Extreme Learning Machine


Field of view


Integrated discrimination improvement


International Society of Urological Pathology


Least absolute shrinkage and selection operator


Machine learning


Multi-layer perceptron


Multi-parametric magnetic resonance imaging


Magnetic resonance imaging


Negative probability value


Prostate cancer


Prostate Imaging Reporting and Data System


Positive probability value


Prostate specific antigen


Receiver operating characteristic curve


Support Vector Machine


T2 weighted imaging


Echo time


Transfer learning nomogram


Transfer learning


Transfer learning signature


Repetition time


Whole slide images


  1. Sung H, Ferlay J, Siegel RL, Laversanne M, Soerjomataram I, Jemal A, Bray F. Global Cancer statistics 2020: GLOBOCAN estimates of incidence and Mortality Worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2021;71(3):209–49.

    Article  PubMed  Google Scholar 

  2. Qiu H, Cao S, Xu R. Cancer incidence, mortality, and burden in China: a time-trend analysis and comparison with the United States and United Kingdom based on the global epidemiological data released in 2020. Cancer Commun (Lond). 2021;41(10):1037–48.

    Article  PubMed  Google Scholar 

  3. Etzioni R, Gulati R, Cooperberg MR, Penson DM, Weiss NS, Thompson IM. Limitations of basing screening policies on screening trials: the US Preventive services Task force and Prostate Cancer screening. Med Care. 2013;51(4):295–300.

    Article  PubMed  PubMed Central  Google Scholar 

  4. Naji L, Randhawa H, Sohani Z, Dennis B, Lautenbach D, Kavanagh O, Bawor M, Banfield L, Profetto J. Digital rectal examination for Prostate Cancer screening in primary care: a systematic review and Meta-analysis. Ann Fam Med. 2018;16(2):149–54.

    Article  PubMed  PubMed Central  Google Scholar 

  5. Eklund M, Jaderling F, Discacciati A, Bergman M, Annerstedt M, Aly M, Glaessgen A, Carlsson S, Gronberg H, Nordstrom T, et al. MRI-Targeted or standard biopsy in Prostate Cancer screening. N Engl J Med. 2021;385(10):908–20.

    Article  PubMed  Google Scholar 

  6. Rouvière O, Puech P, Renard-Penna R, Claudon M, Roy C, Mège-Lechevallier F, Decaussin-Petrucci M, Dubreuil-Chambardel M, Magaud L, Remontet L, et al. Use of prostate systematic and targeted biopsy on the basis of multiparametric MRI in biopsy-naive patients (MRI-FIRST): a prospective, multicentre, paired diagnostic study. Lancet Oncol. 2019;20(1):100–9.

    Article  PubMed  Google Scholar 

  7. EAU-EANM-ESTRO-ESUR-SIOG, Guidelines on Prostate Cancer. (2023 Version) . Accessed 12 Oct 2023.

  8. NCCN Clinical Practice Guidelines in. Oncology-Prostate Cancer Early Detection (2021 Version 2). Accessed 16 Sept 2021.

  9. Turkbey B, Rosenkrantz AB, Haider MA, Padhani AR, Villeirs G, Macura KJ, Tempany CM, Choyke PL, Cornud F, Margolis DJ, et al. Prostate imaging reporting and Data System Version 2.1: 2019 update of prostate imaging reporting and Data System Version 2. Eur Urol. 2019;76(3):340–51.

    Article  PubMed  Google Scholar 

  10. Sathianathen NJ, Omer A, Harriss E, Davies L, Kasivisvanathan V, Punwani S, Moore CM, Kastner C, Barrett T, Van Den Bergh RC, et al. Negative Predictive Value of Multiparametric Magnetic Resonance Imaging in the detection of clinically significant Prostate Cancer in the prostate imaging reporting and Data System era: a systematic review and Meta-analysis. Eur Urol. 2020;78(3):402–14.

    Article  PubMed  Google Scholar 

  11. Park KJ, Choi SH, Kim MH, Kim JK, Jeong IG. Performance of prostate imaging reporting and Data System Version 2.1 for diagnosis of Prostate Cancer: a systematic review and Meta-analysis. J Magn Reson Imaging. 2021;54(1):103–12.

    Article  PubMed  Google Scholar 

  12. De Visschere PJ, Vral A, Perletti G, Pattyn E, Praet M, Magri V, Villeirs GM. Multiparametric magnetic resonance imaging characteristics of normal, benign and malignant conditions in the prostate. Eur Radiol. 2017;27(5):2095–109.

    Article  PubMed  Google Scholar 

  13. Panebianco V, Giganti F, Kitzing YX, Cornud F, Campa R, De Rubeis G, Ciardi A, Catalano C, Villeirs G. An update of pitfalls in prostate mpMRI: a practical approach through the lens of PI-RADS v. 2 guidelines. Insights Imaging. 2018;9(1):87–101.

    Article  PubMed  Google Scholar 

  14. Westphalen AC, McCulloch CE, Anaokar JM, Arora S, Barashi NS, Barentsz JO, Bathala TK, Bittencourt LK, Booker MT, Braxton VG, et al. Variability of the positive predictive value of PI-RADS for prostate MRI across 26 centers: experience of the society of abdominal radiology Prostate Cancer Disease-focused panel. Radiology. 2020;296(1):76–84.

    Article  PubMed  Google Scholar 

  15. Sonn GA, Fan RE, Ghanouni P, Wang NN, Brooks JD, Loening AM, Daniel BL, To’o KJ, Thong AE, Leppert JT. Prostate magnetic resonance imaging interpretation varies substantially across radiologists. Eur Urol Focus. 2019;5(4):592–9.

    Article  PubMed  Google Scholar 

  16. Choy G, Khalilzadeh O, Michalski M, Do S, Samir AE, Pianykh OS, Geis JR, Pandharipande PV, Brink JA, Dreyer KJ. Current applications and future impact of machine learning in Radiology. Radiology. 2018;288(2):318–28.

    Article  PubMed  Google Scholar 

  17. Castaldo R, Cavaliere C, Soricelli A, Salvatore M, Pecchia L, Franzese M. Radiomic and genomic machine learning method performance for Prostate Cancer diagnosis: systematic literature review. J Med Internet Res. 2021;23(4):e22394.

    Article  PubMed  PubMed Central  Google Scholar 

  18. Deniffel D, Abraham N, Namdar K, Dong X, Salinas E, Milot L, Khalvati F, Haider MA. Using decision curve analysis to benchmark performance of a magnetic resonance imaging-based deep learning model for Prostate cancer risk assessment. Eur Radiol. 2020;30(12):6867–76.

    Article  PubMed  Google Scholar 

  19. Zhong X, Cao R, Shakeri S, Scalzo F, Lee Y, Enzmann DR, Wu HH, Raman SS, Sung K. Deep transfer learning-based Prostate cancer classification using 3 Tesla multi-parametric MRI. Abdom Radiol. 2018;44(6):2030–9.

    Article  Google Scholar 

  20. Youn SY, Choi MH, Kim DH, Lee YJ, Huisman H, Johnson E, Penzkofer T, Shabunin I, Winkel DJ, Xing P, et al. Detection and PI-RADS classification of focal lesions in prostate MRI: performance comparison between a deep learning-based algorithm (DLA) and radiologists with various levels of experience. Eur J Radiol. 2021;142:109894.

    Article  PubMed  Google Scholar 

  21. Pan SJ, Yang Q. A survey on transfer learning. IEEE Trans Knowl Data Eng. 2010;22(10):1345–59.

    Article  Google Scholar 

  22. Tan C, Sun F, Kong T, Zhang W, Yang C, Liu C. A Survey on Deep Transfer Learning. In: Artificial Neural Networks and Machine Learning – ICANN 2018 edn.; 2018: 270–279.

  23. Abbasi AA, Hussain L, Awan IA, Abbasi I, Majid A, Nadeem MSA, Chaudhary QA. Detecting Prostate cancer using deep learning convolution neural network with transfer learning approach. Cogn Neurodyn. 2020;14(4):523–33.

    Article  PubMed  PubMed Central  Google Scholar 

  24. de Rooij M, Israël B, Tummers M, Ahmed HU, Barrett T, Giganti F, Hamm B, Løgager V, Padhani A, Panebianco V, et al. ESUR/ESUI consensus statements on multi-parametric MRI for the detection of clinically significant Prostate cancer: quality requirements for image acquisition, interpretation and radiologists’ training. Eur Radiol. 2020;30(10):5404–16.

    Article  PubMed  PubMed Central  Google Scholar 

  25. Humphrey PA, Moch H, Cubilla AL, Ulbright TM, Reuter VE. The 2016 WHO classification of tumours of the urinary system and male genital organs—part B: prostate and bladder tumours. Eur Urol. 2016;70(1):106–19.

    Article  PubMed  Google Scholar 

  26. Tolstikhin IO, Houlsby N, Kolesnikov A, Beyer L, Zhai X, Unterthiner T, Yung J, Steiner A, Keysers D, Uszkoreit J. Mlp-mixer: an all-mlp architecture for vision. Adv Neural Inf Process Syst. 2021;34:24261–72.

    Google Scholar 

  27. The Cancer Genome Atlas. Accessed 10 May 2022.

  28. Feng B, Chen X, Chen Y, Liu K, Li K, Liu X, Yao N, Li Z, Li R, Zhang C, et al. Radiomics nomogram for preoperative differentiation of lung tuberculoma from adenocarcinoma in solitary pulmonary solid nodule. Eur J Radiol. 2020;128:109022.

    Article  PubMed  Google Scholar 

  29. Feng B, Chen X, Chen Y, Yu T, Duan X, Liu K, Li K, Liu Z, Lin H, Li S et al. Identifying Solitary Granulomatous nodules from Solid Lung Adenocarcinoma: exploring robust image features with cross-domain transfer learning. Cancers (Basel). 2023;15(3):892.

  30. Alberts AR, Roobol MJ, Verbeek JFM, Schoots IG, Chiu PK, Osses DF, Tijsterman JD, Beerlage HP, Mannaerts CK, Schimmoller L, et al. Prediction of high-grade Prostate Cancer following multiparametric magnetic resonance imaging: improving the Rotterdam European Randomized Study of screening for Prostate Cancer risk calculators. Eur Urol. 2019;75(2):310–8.

    Article  PubMed  Google Scholar 

  31. Mehralivand S, Shih JH, Rais-Bahrami S, Oto A, Bednarova S, Nix JW, Thomas JV, Gordetsky JB, Gaur S, Harmon SA, et al. A magnetic resonance imaging-based prediction model for prostate biopsy risk stratification. JAMA Oncol. 2018;4(5):678–85.

    Article  PubMed  PubMed Central  Google Scholar 

  32. Lerner LB, McVary KT, Barry MJ, Bixler BR, Dahm P, Das AK, Gandhi MC, Kaplan SA, Kohler TS, Martin L, et al. Management of lower urinary tract symptoms attributed to Benign Prostatic Hyperplasia: AUA GUIDELINE PART I-Initial work-up and Medical Management. J Urol. 2021;206(4):806–817.

    Article  PubMed  Google Scholar 

  33. Kwak JT, Sankineni S, Xu S, Turkbey B, Choyke PL, Pinto PA, Moreno V, Merino M, Wood BJ. Prostate Cancer: a correlative study of Multiparametric MR Imaging and Digital Histopathology. Radiology. 2017;285(1):147–56.

    Article  PubMed  Google Scholar 

  34. Raghu M, Zhang C, Kleinberg J, Bengio S. Transfusion: understanding transfer learning for medical imaging. Adv Neural Inf Process Syst. 2019;32.

  35. Feng B, Huang L, Liu Y, Chen Y, Zhou H, Yu T, Xue H, Chen Q, Zhou T, Kuang Q et al. A transfer learning Radiomics Nomogram for Preoperative Prediction of Borrmann Type IV gastric Cancer from primary gastric Lymphoma. Front Oncol. 2022;11:802205.

Download references


Not applicable.


This work was supported by the National Natural Science Foundation of China [81960324, 62176104] and the Natural Science Foundation of Guangxi Province [2021GXNSFAA075037], the funding bodies played no role in the design of the study and collection, analysis, interpretation of data, and in writing the manuscript.

Author information

Authors and Affiliations



JC contributed to the design of the study, the acquisition and interpretation of data and had drafted the manuscript. BF contributed to the conception of the study, the analysis of data and had revised the manuscript. MH contributed to the acquisition and interpretation of data. FH contributed to the analysis and interpretation of data and had drafted the manuscript. YC contributed to the analysis of data and had revised the manuscript. XM contributed to the acquisition and interpretation of data. WL contributed to the conception of the study, the interpretation of data and revised the manuscript. All authors reviewed the manuscript.

Corresponding author

Correspondence to Wansheng Long.

Ethics declarations

Ethics approval and consent to participate

This retrospective study was approved by the Institutional Review Board of The Jiangmen Central Hospital (Approval No.39; 2023). This study was carried out in accordance with the Declaration of Helsinki. The informed consent was waived by the Institutional Review Board of The Jiangmen Central Hospital.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.

Additional information

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit The Creative Commons Public Domain Dedication waiver ( applies to the data made available in this article, unless otherwise stated in a credit line to the data.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Chen, J., Feng, B., Hu, M. et al. A transfer learning nomogram for predicting prostate cancer and benign conditions on MRI. BMC Med Imaging 23, 200 (2023).

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: