Navigating the gray zone: Machine learning can differentiate malignancy in PI-RADS 3 lesions.

INTRODUCTION

The objective of this study is to predict the probability of prostate cancer in PI-RADS 3 lesions using machine learning methods that incorporate clinical and mpMRI parameters.

METHODS

The study included patients who had PI-RADS 3 lesions detected on mpMRI and underwent fusion biopsy between January 2020 and January 2024. Radiological parameters (Apparent diffusion coefficient (ADC), tumour ADC/contralateral ADC ratio, Ktrans value, periprostatic adipose tissue thickness, lesion size, prostate volume) and clinical parameters (age, body mass index, total prostate specific antigen, free PSA, PSA density, systemic inflammatory index, neutrophil-lymphocyte ratio [NLR], platelet lymphocyte ratio, lymphocyte monocyte ratio) were documented. The probability of prostate cancer prediction in PI-RADS 3 lesions was calculated using 6 different machine-learning models, with the input parameters being the aforementioned variables.

RESULTS

Of the 235 participants in the trial, 61 had malignant fusion biopsy pathology and 174 had benign pathology. Among 6 different machine learning algorithms, the random forest model had the highest accuracy (0.86±0.04; 95% CI 0.85-0.87), F1 score (0.91±0.03; 95% CI 0.91-0.92) and AUC value (0.92±0.06; 95% CI 0.88-0.90). In SHAP analysis based on random forest model, tumour ADC, tumour ADC/contralateral ADC ratio and PSA density were the 3 most successful parameters in predicting malignancy. On the other hand, systemic inflammatory index and neutrophil lymphocyte ratio showed higher accuracy in predicting malignancy than total PSA, age, free PSA/total PSA and lesion size in SHAP analysis.

CONCLUSION

Among the machine learning models we developed, especially the random forest model can predict malignancy in PI-RADS 3 lesions and prevent unnecessary biopsy. This model can be used in clinical practice with multicentre studies including more patients.