Magnetic Resonance Imaging-Based Predictive Models for Clinically Significant Prostate Cancer: A Systematic Review

Adaptive fusion of dual-view for grading prostate cancer

Accurate preoperative grading of prostate cancer is crucial for assisted diagnosis. Multi-parametric magnetic resonance imaging (MRI) is a commonly used non-invasive approach, however, the interpretation of MRI images is still subject to significant subjectivity due to variations in physicians' expertise and experience. To achieve accurate, non-invasive, and efficient grading of prostate cancer, this paper proposes a deep learning method that adaptively fuses dual-view MRI images. Specifically, a dual-view adaptive fusion model is designed. The model employs encoders to extract embedded features from two MRI sequences: T2-weighted imaging (T2WI) and apparent diffusion coefficient (ADC). The model reconstructs the original input images using the embedded features and adopts a cross-embedding fusion module to adaptively fuse the embedded features from the two views. Adaptive fusion refers to dynamically adjusting the fusion weights of the features from the two views according to different input samples, thereby fully utilizing complementary information. Furthermore, the model adaptively weights the prediction results from the two views based on uncertainty estimation, further enhancing the grading performance. To verify the importance of effective multi-view fusion for prostate cancer grading, extensive experiments are designed. The experiments evaluate the performance of single-view models, dual-view models, and state-of-the-art multi-view fusion algorithms. The results demonstrate that the proposed dual-view adaptive fusion method achieves the best grading performance, confirming its effectiveness for assisted grading diagnosis of prostate cancer. This study provides a novel deep learning solution for preoperative grading of prostate cancer, which has the potential to assist clinical physicians in making more accurate diagnostic decisions and has significant clinical application value.

A Novel Machine Learning-Based Predictive Model of Clinically Significant Prostate Cancer and Online Risk Calculator

OBJECTIVES

To create a machine learning predictive model combining PI-RADS score, PSA density, and clinical variables to predict clinically significant prostate cancer (csPCa).

METHODS

We evaluated a cohort of patients who underwent prostate biopsy for suspected prostate cancer (PCa) in New Zealand, Australia, and Switzerland. We collected data on age, body mass index (BMI), PSA level, prostate volume, PSA density (PSAD), PI-RADS scores, previous biopsy, and corresponding histology results. The dataset was divided into derivation (training) and validation (test) sets using random splits. An independent dataset was obtained from the Harvard Dataverse for external validation. A cohort of 1272 patients was analyzed. We fitted a Lasso model, XGBoost, and LightGBM to the training set and assessed their accuracy.

RESULTS

All models demonstrated ROC AUC values ranging from 0.830 to 0.851. LightGBM was considered the superior model, with an ROC of 0.851 [95%CI: 0.804 - 0.897] in the test set and 0.818 [95% CI: 0.798 - 0.831] in the external dataset. The most important variable was PI-RADS, followed by PSA density, history of previous biopsy, age, and BMI.

CONCLUSIONS

We developed a predictive model for detecting csPCa that exhibited a high ROC-AUC value for internal and external validations. This suggests that the integration of the clinical parameters outperformed each individual predictor. Additionally, the model demonstrated good calibration metrics, indicative of a more balanced model than the existing models.

Assessment of a fully-automated diagnostic AI software in prostate MRI: Clinical evaluation and histopathological correlation

OBJECTIVE

This study aims to evaluate the diagnostic performance of a commercial, fully-automated, artificial intelligence (AI) driven software tool in identifying and grading prostate lesions in prostate MRI, using histopathological findings as the reference standard, while contextualizing its performance within the framework of PI-RADS v2.1 criteria.

MATERIAL AND METHODS

This retrospective study analyzed 123 patients who underwent multiparametric prostate MRI followed by systematic and targeted biopsies. MRI protocols adhered to international guidelines and included T2-weighted, diffusion-weighted, T1-weighted, and dynamic contrast-enhanced imaging. The AI software tool mdprostate was integrated into the Picture Archiving and Communication System to automatically segment the prostate, calculate prostate volume, and classify lesions according to PI-RADS scores using biparametric T2-weighted and diffusion-weighted imaging. Histopathological analysis of biopsy cores served as the reference standard. Diagnostic performance metrics including sensitivity, specificity, positive and negative predictive value (PPV, NPV), and area under the ROC curve (AUC) were calculated.

RESULTS

mdprostate demonstrated 100 % sensitivity at a PI-RADS ≥ 2 cutoff, effectively ruling out both clinically significant and non-significant prostate cancers for lesions remaining below this threshold. For detecting clinically significant prostate cancer (csPCa) using a PI-RADS ≥ 4 cutoff, mdprostate achieved a sensitivity of 85.5 % and a specificity of 63.2 %. The AUC for detecting cancers of any grade was 0.803. The performance metrics of mdprostate were comparable to those reported in two meta-analyses of PI-RADS v2.1, with no significant differences in sensitivity and specificity (p > 0.05).

CONCLUSION

The evaluated AI tool demonstrated high diagnostic performance in identifying and grading prostate lesions, with results comparable to those reported in meta-analyses of expert readers using PI-RADS v2.1. Its ability to standardize evaluations and potentially reduce variability underscores its potential as a valuable adjunct in the prostate cancer diagnostic pathway. The high accuracy of mdprostate, particularly in ruling out prostate cancers, highlights its clinical utility by reducing workload and enhancing patient outcomes.

Validation of the Barcelona-MRI predictive model when PI-RADS v2.1 is used with trans-perineal prostate biopsies

PURPOSE

To validate the Barcelona magnetic resonance imaging predictive model (BCN-MRI PM) in men with pre-biopsy multiparametric MRI (mpMRI) reported with the Prostate Imaging Reporting and Data System (PI-RADS) v2.1, followed by transrectal and transperineal prostate biopsies.

MATERIALS AND METHODS

Prospective analysis of 3,264 men with PSA >3.0 ng/mL and/or abnormal digital rectal examination who were referred to ten participant centers in the csPCa early detection program of Catalonia (Spain), between 2021 and 2023. MpMRI was reported with the PI-RADS v2.1, and 2- to 4-core MRI-transrectal ultrasound (TRUS) fusion-targeted biopsy of suspected lesions and/or 12-core systematic biopsy were conducted. 2,295 (70.3%) individuals were referred to six centers for transrectal prostate biopsies, while 969 (39.7%) were referred to four centers for transperineal prostate biopsies. CsPCa was classified whenever the International Society of Urologic Pathology grade group was 2 or higher.

RESULTS

CsPCa was detected in 41% of transrectal prostate biopsies and in 45.9% of transperineal prostate biopsies (p < 0.016). Both BCN-MRI PM calibration curves were within the ideal correlation between predicted and observed csPCa. Areas under the curve and 95% confidence intervals were 0.847 (0.830-0.857) and 0.830 (0.823-0.855), respectively (p = 0.346). Specificities corresponding to 95% sensitivity were 37.6 and 36.8%, respectively (p = 0.387). The Net benefit of the BCN-MRI PM was similar with both biopsy methods.

CONCLUSIONS

The BCN-MRI PM has been successfully validated when mpMRI was reported with the PI-RADS v2.1 and prostate biopsies were conducted via the transrectal and transperineal route.

External validation of the barcelona magnetic resonance imaging predictive model for detecting significant prostate cancer including men receiving 5-alpha reductase inhibitors

PURPOSE

To validate the Barcelona-magnetic resonance imaging predictive model (BCN-MRI PM) for clinically significant prostate cancer (csPCa) in Catalonia, a Spanish region with 7.9 million inhabitants. Additionally, the BCN-MRI PM is validated in men receiving 5-alpha reductase inhibitors (5-ARI).

MATERIALS AND METHODS

A population of 2,212 men with prostate-specific antigen serum level > 3.0 ng/ml and/or a suspicious digital rectal examination who underwent multiparametric MRI and targeted and/or systematic biopsies in the year 2022, at ten participant centers of the Catalonian csPCa early detection program, were selected. 120 individuals (5.7%) were identified as receiving 5-ARI treatment for longer than a year. The risk of csPCa was retrospectively assessed with the Barcelona-risk calculator 2 (BCN-RC 2). Men undergoing 5-ARI treatment for less than a year were excluded. CsPCa was defined when the grade group was ≥ 2.

RESULTS

The area under the curve of the BCN-MRI PM in 5-ARI naïve men was 0.824 (95% CI 0.783-0.842) and 0.849 (0.806-0.916) in those receiving 5-ARI treatment, p 0.475. Specificities at 100, 97.5, and 95% sensitivity thresholds were to 2.7, 29.3, and 39% in 5-ARI naïve men, while 43.5, 46.4, and 47.8%, respectively in 5-ARI users. The application of BCN-MRI PM would result in a reduction of 23.8% of prostate biopsies missing 5% of csPCa in 5-ARI naïve men, while reducing 25% of prostate biopsies without missing csPCa in 5-ARI users.

CONCLUSIONS

The BCN-MRI PM has achieved successful validation in Catalonia and, notably, for the first time, in men undergoing 5-ARI treatment.

External validation and comparison of magnetic resonance imaging-based risk prediction models for prostate biopsy stratification

PURPOSE

Magnetic resonance imaging (MRI) is a promising tool for risk assessment, potentially reducing the burden of unnecessary prostate biopsies. Risk prediction models that incorporate MRI data have gained attention, but their external validation and comparison are essential for guiding clinical practice. The aim is to externally validate and compare risk prediction models for the diagnosis of clinically significant prostate cancer (csPCa).

METHODS

A cohort of 4606 patients across fifteen European tertiary referral centers were identified from a prospective maintained database between January 2016 and April 2023. Transrectal or transperineal image-fusion MRI-targeted and systematic biopsies for PI-RADS score of ≥ 3 or ≥ 2 depending on patient characteristics and physician preferences. Probabilities for csPCa, defined as International Society of Urological Pathology (ISUP) grade ≥ 2, were calculated for each patients using eight models. Performance was characterized by area under the receiver operating characteristic curve (AUC), calibration, and net benefit. Subgroup analyses were performed across various clinically relevant subgroups.

RESULTS

Overall, csPCa was detected in 2154 (47%) patients. The models exhibited satisfactory performance, demonstrating good discrimination (AUC ranging from 0.75 to 0.78, p < 0.001), adequate calibration, and high net benefit. The model described by Alberts showed the highest clinical utility for threshold probabilities between 10 and 20%. Subgroup analyses highlighted variations in models' performance, particularly when stratified according to PSA level, biopsy technique and PI-RADS version.

CONCLUSIONS

We report a comprehensive external validation of risk prediction models for csPCa diagnosis in patients who underwent MRI-targeted and systematic biopsies. The model by Alberts demonstrated superior clinical utility and should be favored when determining the need for a prostate biopsy.

Reducing the demand for magnetic resonance imaging scans and prostate biopsies during the early detection of clinically significant prostate cancer: Applying the Barcelona risk-stratified pathway in Catalonia

PURPOSE

To analyze the reduction in multiparametric magnetic resonance imaging (mpMRI) demand and prostate biopsies after the hypothetical implementation of the Barcelona risk-stratified pathway (BCN-RSP) in a population of the clinically significant prostate cancer (csCaP) early detection program in Catalonia.

MATERIALS AND METHODS

A retrospective comparation between the hypothetical application of the BCN-RSP and the current pathway, which relied on pre-biopsy mpMRI and targeted and/or systematic biopsies, was conducted. The BCN-RSP stratify men with suspected CaP based on a prostate specific antigen (PSA) level >10 ng/ml and a suspicious rectal examination (DRE), and the Barcelona-risk calculator 1 (BCN-RC1) to avoid mpMRI scans. Subsequently, candidates for prostate biopsy following mpMRI are selected based on the BCN-RC2. This comparison involved 3,557 men with serum PSA levels > 3.0 ng/ml and/or suspicious DRE. The population was recruited prospectively in 10 centers from January 2021 and December 2022. CsCaP was defined when grade group ≥ 2.

RESULTS

CsCaP was detected in 1,249 men (35.1%) and insignificant CaP was overdeteced in 498 (14%). The BCN-RSP would have avoid 705 mpMRI scans (19.8%), and 697 prostate biopsies (19.6%), while 61 csCaP (4.9%) would have been undetected. The overdetection of insignificant CaP would have decrease in 130 cases (26.1%), and the performance of prostate biopsy for csCaP detection would have increase to 41.5%.

CONCLUSION

The application of the BCN-RSP would reduce the demand for mpMRI scans and prostate biopsies by one fifth while less than 5% of csCaP would remain undetected. The overdetection of insignificant CaP would decrease by more than one quarter and the performance of prostate biopsy for csCaP detection would increase to higher than 40%.

Comparison of Magnetic Resonance Imaging-Based Risk Calculators to Predict Prostate Cancer Risk

Importance

Magnetic resonance imaging (MRI)-based risk calculators can replace or augment traditional prostate cancer (PCa) risk prediction tools. However, few data are available comparing performance of different MRI-based risk calculators in external cohorts across different countries or screening paradigms.

Objective

To externally validate and compare MRI-based PCa risk calculators (Prospective Loyola University Multiparametric MRI [PLUM], UCLA [University of California, Los Angeles]-Cornell, Van Leeuwen, and Rotterdam Prostate Cancer Risk Calculator-MRI [RPCRC-MRI]) in cohorts from Europe and North America.

Design, Setting, and Participants

This multi-institutional, external validation diagnostic study of 3 unique cohorts was performed from January 1, 2015, to December 31, 2022. Two cohorts from Europe and North America used MRI before biopsy, while a third cohort used an advanced serum biomarker, the Prostate Health Index (PHI), before MRI or biopsy. Participants included adult men without a PCa diagnosis receiving MRI before prostate biopsy.

Interventions

Prostate MRI followed by prostate biopsy.

Main Outcomes and Measures

The primary outcome was diagnosis of clinically significant PCa (grade group ≥2). Receiver operating characteristics for area under the curve (AUC) estimates, calibration plots, and decision curve analysis were evaluated.

Results

A total of 2181 patients across the 3 cohorts were included, with a median age of 65 (IQR, 58-70) years and a median prostate-specific antigen level of 5.92 (IQR, 4.32-8.94) ng/mL. All models had good diagnostic discrimination in the European cohort, with AUCs of 0.90 for the PLUM (95% CI, 0.86-0.93), UCLA-Cornell (95% CI, 0.86-0.93), Van Leeuwen (95% CI, 0.87-0.93), and RPCRC-MRI (95% CI, 0.86-0.93) models. All models had good discrimination in the North American cohort, with an AUC of 0.85 (95% CI, 0.80-0.89) for PLUM and AUCs of 0.83 for the UCLA-Cornell (95% CI, 0.80-0.88), Van Leeuwen (95% CI, 0.79-0.88), and RPCRC-MRI (95% CI, 0.78-0.87) models, with somewhat better calibration for the RPCRC-MRI and PLUM models. In the PHI cohort, all models were prone to underestimate clinically significant PCa risk, with best calibration and discrimination for the UCLA-Cornell (AUC, 0.83 [95% CI, 0.81-0.85]) model, followed by the PLUM model (AUC, 0.82 [95% CI, 0.80-0.84]). The Van Leeuwen model was poorly calibrated in all 3 cohorts. On decision curve analysis, all models provided similar net benefit in the European cohort, with higher benefit for the PLUM and RPCRC-MRI models at a threshold greater than 22% in the North American cohort. The UCLA-Cornell model demonstrated highest net benefit in the PHI cohort.

Conclusions and Relevance

In this external validation study of patients receiving MRI and prostate biopsy, the results support the use of the PLUM or RPCRC-MRI models in MRI-based screening pathways regardless of European or North American setting. However, tools specific to screening pathways incorporating advanced biomarkers as reflex tests are needed due to underprediction.

Investigating Efficient Risk-Stratified Pathways for the Early Detection of Clinically Significant Prostate Cancer

Risk-stratified pathways (RSPs) are recommended by the European Association of Uro-logy (EAU) to improve the early detection of clinically significant prostate cancer (csPCa). RSPs can reduce magnetic resonance imaging (MRI) demand, prostate biopsies, and the over-detection of insignificant PCa (iPCa). Our goal is to analyze the efficacy and cost-effectiveness of several RSPs by using sequential stratifications from the serum prostate-specific antigen level and digital rectal examination, the Barcelona risk calculators (BCN-RCs), MRI, and Proclarix™. In a cohort of 567 men with a serum PSA level above 3.0 ng/mL who underwent multiparametric MRI (mpMRI) and targeted and/or systematic biopsies, the risk of csPCa was retrospectively assessed using Proclarix™ and BCN-RCs 1 and 2. Six RSPs were compared with those recommended by the EAU that, stratifying men from MRI, avoided 16.7% of prostate biopsies with a prostate imaging-reporting and data system score of <3, with 2.6% of csPCa cases remaining undetected. The most effective RSP avoided mpMRI exams in men with a serum PSA level of >10 ng/mL and suspicious DRE, following stratifications from BCN-RC 1, mpMRI, and Proclarix™. The demand for mpMRI decreased by 19.9%, prostate biopsies by 19.8%, and over-detection of iPCa by 22.7%, while 2.6% of csPCa remained undetected as in the recommended RSP. Cost-effectiveness remained when the Proclarix™ price was assumed to be below EUR 200.

A Novel Nomogram to Identify Candidates for Focal Therapy Among Patients with Localized Prostate Cancer Diagnosed via Magnetic Resonance Imaging-Targeted and Systematic Biopsies: A European Multicenter Study

BACKGROUND

Suitable selection criteria for focal therapy (FT) are crucial to achieve success in localized prostate cancer (PCa).

OBJECTIVE

To develop a multivariable model that better delineates eligibility for FT and reduces undertreatment by predicting unfavorable disease at radical prostatectomy (RP).

DESIGN, SETTING, AND PARTICIPANTS

Data were retrospectively collected from a prospective European multicenter cohort of 767 patients who underwent magnetic resonance imaging (MRI)-targeted and systematic biopsies followed by RP in eight referral centers between 2016 and 2021. The Imperial College of London eligibility criteria for FT were applied: (1) unifocal MRI lesion with Prostate Imaging-Reporting and Data System score of 3-5; (2) prostate-specific antigen (PSA) ≤20 ng/ml; (3) cT2-3a stage on MRI; and (4) International Society of Urological Pathology grade group (GG) 1 and ≥6 mm or GG 2-3. A total of 334 patients were included in the final analysis.

OUTCOME MEASUREMENTS AND STATISTICAL ANALYSIS

The primary outcome was unfavorable disease at RP, defined as GG ≥4, and/or lymph node invasion, and/or seminal vesicle invasion, and/or contralateral clinically significant PCa. Logistic regression was used to assess predictors of unfavorable disease. The performance of the models including clinical, MRI, and biopsy information was evaluated using the area under the receiver operating characteristic curve (AUC), calibration plots, and decision curve analysis. A coefficient-based nomogram was developed and internally validated.

RESULTS AND LIMITATIONS

Overall, 43 patients (13%) had unfavorable disease on RP pathology. The model including PSA, clinical stage on digital rectal examination, and maximum lesion diameter on MRI had an AUC of 73% on internal validation and formed the basis of the nomogram. Addition of other MRI or biopsy information did not significantly improve the model performance. Using a cutoff of 25%, the proportion of patients eligible for FT was 89% at the cost of missing 30 patients (10%) with unfavorable disease. External validation is required before the nomogram can be used in clinical practice.

CONCLUSIONS

We report the first nomogram that improves selection criteria for FT and limits the risk of undertreatment.

PATIENT SUMMARY

We conducted a study to develop a better way of selecting patients for focal therapy for localized prostate cancer. A novel predictive tool was developed using the prostate-specific antigen (PSA) level measured before biopsy, tumor stage assessed via digital rectal examination, and lesion size on magnetic resonance imaging (MRI) scans. This tool improves the prediction of unfavorable disease and may reduce the risk of undertreatment of localized prostate cancer when using focal therapy.

Development and validation of a prognostic nomogram for gastrointestinal stromal tumors in the postimatinib era: A study based on the SEER database and a Chinese cohort

BACKGROUND

After the standardization, recording and follow-up of imatinib use that significantly prolongs survival of gastrointestinal stromal tumors (GISTs), a comprehensive reassessment of the prognosis of GISTs is necessary and more conductive to treatment options.

METHODS

A total of 2185 GISTs between 2013 and 2016 were obtained from the Surveillance, Epidemiology, and End Results database and comprised our training (n = 1456) and internal validation cohorts (n = 729). The risk factors extracted from univariate and multivariate analyses were used to establish a predictive nomogram. The model was evaluated and tested in the validation cohort internally and in 159 patients with GIST diagnosed between January 2015 and June 2017 in Xijing Hospital externally.

RESULTS

The median OS was 49 months (range, 0-83 months) in the training cohort and 51 months (0-83 months) in the validation cohort. The concordance index (C-index) of the nomogram was 0.777 (95% CI, 0.752-0.802) and 0.7787 (0.7785, bootstrap corrected) in training and internal validation cohorts, respectively, and 0.7613 (0.7579, bootstrap corrected) in the external validation cohort. Receiver operating characteristic curves and calibration curves for 1-, 3-, and 5-year overall survival (OS) showed a high degree of discrimination and calibration. The area under the curve showed that the new model performed better than the TNM staging system. In addition, the model could be dynamically visualized on a webpage.

CONCLUSION

We developed a comprehensive survival prediction model for assessing the 1-, 3- and 5-year OS of patients with GIST in the postimatinib era. This predictive model outperforms the traditional TNM staging system and sheds light on the improvement of the prognostic prediction and the selection of treatment strategies for GISTs.

Comparison of Rotterdam and Barcelona Magnetic Resonance Imaging Risk Calculators for Predicting Clinically Significant Prostate Cancer

Background

Magnetic resonance imaging (MRI)-based risk calculators (MRI-RCs) individualise the likelihood of clinically significant prostate cancer (csPCa) and improve candidate selection for prostate biopsy beyond the Prostate Imaging Reporting and Data System (PI-RADS).

Objective

To compare the Barcelona (BCN) and Rotterdam (ROT) MRI-RCs in an entire population and according to the PI-RADS categories.

Design setting and participants

A prospective comparison of BCN- and ROT-RC in 946 men with suspected prostate cancer in whom systematic biopsy was performed, as well as target biopsies of PI-RADS ≥3 lesions.

Outcome measurements and statistical analysis

Saved biopsies and undetected csPCa (grade group ≥2) were determined.

Results and limitations

The csPCa detection was 40.8%. The median risks of csPCa from BCN- and ROT-RC were, respectively, 67.1% and 25% in men with csPCa, whereas 10.5% and 3% in those without csPCa (p < 0.001). The areas under the curve were 0.856 and 0.844, respectively (p = 0.116). BCN-RC showed a higher net benefit and clinical utility over ROT-RC. Using appropriate thresholds, respectively, 75% and 80% of biopsies were needed to identify 50% of csPCa detected in men with PI-RADS <3, whereas 35% and 21% of biopsies were saved, missing 10% of csPCa detected in men with PI-RADS 3. BCN-RC saved 15% of biopsies, missing 2% of csPCa in men with PI-RADS 4, whereas ROT-RC saved 10%, missing 6%. No RC saved biopsies without missing csPCa in men with PI-RADS 5.

Conclusions

ROT-RC provided a lower and narrower range of csPCa probabilities than BCN-RC. BCN-RC showed a net benefit over ROT-RC in the entire population. However, BCN-RC was useful in men with PI-RADS 3 and 4, whereas ROT-RC was useful only in those with PI-RADS 3. No RC seemed to be helpful in men with negative MRI and PI-RADS 5.

Patient summary

Barcelona risk calculator was more helpful than Rotterdam risk calculator to select candidates for prostate biopsy.

Beyond Multiparametric MRI and towards Radiomics to Detect Prostate Cancer: A Machine Learning Model to Predict Clinically Significant Lesions

The risk of misclassifying clinically significant prostate cancer (csPCa) by multiparametric magnetic resonance imaging is consistent, also using the updated PIRADS score and although different definitions of csPCa, patients with Gleason Grade group (GG) ≥ 3 have a significantly worse prognosis. This study aims to develop a machine learning model predicting csPCa (i.e., any GG ≥ 3 lesion at target biopsy) by mpMRI radiomic features and analyzing similarities between GG groups. One hundred and two patients with 117 PIRADS ≥ 3 lesions at mpMRI underwent target+systematic biopsy, providing histologic diagnosis of PCa, 61 GG < 3 and 56 GG ≥ 3. Features were generated locally from an apparent diffusion coefficient and selected, using the LASSO method and Wilcoxon rank-sum test (p < 0.001), to achieve only four features. After data augmentation, the features were exploited to train a support vector machine classifier, subsequently validated on a test set. To assess the results, Kruskal−Wallis and Wilcoxon rank-sum tests (p < 0.001) and receiver operating characteristic (ROC)-related metrics were used. GG1 and GG2 were equivalent (p = 0.26), whilst clear separations between either GG[1,2] and GG ≥ 3 exist (p < 10−6). On the test set, the area under the curve = 0.88 (95% CI, 0.68−0.94), with positive and negative predictive values being 84%. The features retain a histological interpretation. Our model hints at GG2 being much more similar to GG1 than GG ≥ 3.