Multiparametric Magnetic Resonance Imaging Histology for Assessing Tumor Interstitial Ratio in Prostate Cancer

Objective Multiparameter magnetic resonance imaging (mpMRI) was used to evaluate the tumor interstitial ratio (TSR) of prostate cancer (PCa), and to explore the correlation between the tumor interstitial ratio and the findings of multiparameter magnetic resonance imaging and the correlation between the tumor interstitial ratio and prostate cancer prognosis. Methods General data, preoperative magnetic resonance (MRI) and postoperative pathology of our prostate cancer patients from January 2020 to January 2023 were retrospectively collected. The section with the most interstitial content was selected under 10×10 lens and the area with the most interstitial content was selected, and the selected area was observed by switching the 10×40 lens. 9 images were obtained in 3×3 format, which were colored and quantified using ImageJ to obtain the area ratio of tumor cells (T) and interstitial cells (S), and the TSR was evaluated, and the correlation between general data and TSR was calculated. Using 3D Slicer, T1-weighted images (T1WI), apparent diffusion coefficient (ADC), dynamic enhancement images (DCE) and diffusion-weighted images (DWI) were registered based on T2-weighted imaging (T2WI), and semi-automatic 3D mapping was performed along the lesion on T2-weighted imaging (T2WI). The region of interest (3D-ROI) of T2WI was obtained, and the remaining 4 sequences automatically obtained the region of interest (3D-ROI) after multi-modal learning due to registration, feature extraction was carried out, and structural random forest (SRF) was used for feature selection. Finally, five machine learning algorithms including K-nearest neighbor classification (KNN), random forest , support vector machine (SVM), extreme gradient Boost (XGBoost) and LightGBM (LightGBM), combined with meaningful feature variables, were used to build prediction models, and receiver operating characteristic curves of each model were obtained. The area under the curve (AUC) was calculated, and the accuracy, accuracy, recall, specificity, F1 value, negative predictive value (NPV) and positive predictive value (PPV) of different machine learning algorithms were calculated by means of 50% cross-validation. And calculate the accuracy, precision, recall, specificity, F1 value, negative predictive value (NPV) and positive predictive value (PPV) of each sequence in the optimal machine learning algorithm model. Results The five models were successfully established and the results were relatively consistent, among which LightGBM model showed the most stable performance with an accuracy of 0.766. In addition, the accuracy, recall, F1 value, specificity, negative predictive value (NPV) and positive predictive value (PPV) were 0.784, 0.674, 0.843, 0.725, 0.784 and 0.754, respectively. In LightGBM model, DWI sequence showed the best performance, with accuracy, accuracy, recall, F1 value, specificity, negative predictive value (NPV) and positive predictive value (PPV) being 0.660, 0.657, 0.535, 0.765, 0.590, 0.657 and 0.661, respectively. Conclusion A prostate cancer imaging model based on multi-parameter magnetic resonance has been successfully developed and verified. Preoperative prostate cancer MRI imaging combined with machine learning algorithm can predict the tumor interstitial ratio, and higher interstitial is more aggressive than lower interstitial, and the prognosis is worse, which has scientific significance for promoting the individualized treatment of prostate cancer.