The practical clinical role of machine learning models with different algorithms in predicting prostate cancer local recurrence after radical prostatectomy

BACKGROUND

The detection of local recurrence for prostate cancer (PCa) patients following radical prostatectomy (RP) is challenging and can influence the treatment plan. Our aim was to construct and verify machine learning models with three different algorithms based on post-operative mpMRI for predicting local recurrence of PCa after RP and explore their potential clinical value compared with the Prostate Imaging for Recurrence Reporting (PI-RR) score of expert-level radiologists.

METHODS

A total of 176 patients were retrospectively enrolled and randomly divided into training (n = 123) and testing (n = 53) sets. The PI-RR assessments were performed by two expert-level radiologists with access to the operative histopathological and pre-surgical clinical results. The radiomics models to predict local recurrence were built by utilizing three different algorithms (i.e., support vector machine [SVM], linear discriminant analysis [LDA], and logistic regression-least absolute shrinkage and selection operator [LR-LASSO]). The combined model integrating radiomics features and PI-RR score was developed using the most effective classifier. The classification performances of the proposed models were assessed by receiver operating characteristic (ROC) curve analysis.

RESULTS

There were no significant differences between the training and testing sets concerning age, prostate-specific antigen (PSA), Gleason score, T-stage, seminal vesicle invasion (SVI), perineural invasion (PNI), and positive surgical margins (PSM). The radiomics model based on LR-LASSO exhibited superior performance than other radiomics models, with an AUC of 0.858 in the testing set; the PI-RR yielded an AUC of 0.833, and there was no significant difference between the best radiomics model and the PI-RR score. The combined model achieved the best predictive performance with an AUC of 0.924, and a significant difference was observed between the combined model and PI-RR score.

CONCLUSIONS

Our radiomics model is an effective tool to predict PCa local recurrence after RP. By integrating radiomics features with the PI-RR score, our combined model exhibited significantly better predictive performance of local recurrence than expert-level radiologists' PI-RR assessment.

Machine learning-based radiomics model to predict benign and malignant PI-RADS v2.1 category 3 lesions: a retrospective multi-center study

Purpose

To develop machine learning-based radiomics models derive from different MRI sequences for distinction between benign and malignant PI-RADS 3 lesions before intervention, and to cross-institution validate the generalization ability of the models.

Methods

The pre-biopsy MRI datas of 463 patients classified as PI-RADS 3 lesions were collected from 4 medical institutions retrospectively. 2347 radiomics features were extracted from the VOI of T2WI, DWI and ADC images. The ANOVA feature ranking method and support vector machine classifier were used to construct 3 single-sequence models and 1 integrated model combined with the features of three sequences. All the models were established in the training set and independently verified in the internal test and external validation set. The AUC was used to compared the predictive performance of PSAD with each model. Hosmer–lemeshow test was used to evaluate the degree of fitting between prediction probability and pathological results. Non-inferiority test was used to check generalization performance of the integrated model.

Results

The difference of PSAD between PCa and benign lesions was statistically significant (P = 0.006), with the mean AUC of 0.701 for predicting clinically significant prostate cancer (internal test AUC = 0.709 vs. external validation AUC = 0.692, P = 0.013) and 0.630 for predicting all cancer (internal test AUC = 0.637 vs. external validation AUC = 0.623, P = 0.036). T2WI-model with the mean AUC of 0.717 for predicting csPCa (internal test AUC = 0.738 vs. external validation AUC = 0.695, P = 0.264) and 0.634 for predicting all cancer (internal test AUC = 0.678 vs. external validation AUC = 0.589, P = 0.547). DWI-model with the mean AUC of 0.658 for predicting csPCa (internal test AUC = 0.635 vs. external validation AUC = 0.681, P = 0.086) and 0.655 for predicting all cancer (internal test AUC = 0.712 vs. external validation AUC = 0.598, P = 0.437). ADC-model with the mean AUC of 0.746 for predicting csPCa (internal test AUC = 0.767 vs. external validation AUC = 0.724, P = 0.269) and 0.645 for predicting all cancer (internal test AUC = 0.650 vs. external validation AUC = 0.640, P = 0.848). Integrated model with the mean AUC of 0.803 for predicting csPCa (internal test AUC = 0.804 vs. external validation AUC = 0.801, P = 0.019) and 0.778 for predicting all cancer (internal test AUC = 0.801 vs. external validation AUC = 0.754, P = 0.047).

Conclusions

The radiomics model based on machine learning has the potential to be a non-invasive tool to distinguish cancerous, noncancerous and csPCa in PI-RADS 3 lesions, and has relatively high generalization ability between different date set.