An Apparent Diffusion Coefficient-Based Machine Learning Model Can Improve Prostate Cancer Detection in the Grey Area of the Prostate Imaging Reporting and Data System Category 3: A Single-Centre Experience

Q: How can radiomics improve the diagnosis of prostate cancer in PI-RADS 3 lesions?

Radiomics, which involves the use of machine learning and artificial intelligence strategies to analyze radiological images, can enrich the information retrieved by visual analysis. By generating quantitative measurements called radiomic features (RFs), radiomics can potentially diagnose benign and malignant lesions in PI-RADS 3 lesions. This approach can improve the predictive capability of the PI-RADS 3 score, allowing for more accurate diagnosis of prostate cancer. Previous studies have shown promising results in using radiomics for this purpose, and the aim of this study is to develop a machine learning model that predicts prostate cancer in a selected cohort of equivocal PI-RADS score 3 lesions, further improving the current state of the art through an automated pipeline of image processing and automatic feature generation.

Q: What were the inclusion criteria for the study population?

The inclusion criteria for the study population were as follows: (1) MRI-TRUS fusion targeted biopsy (fusion-TB) only performed at the Radiology Unit; (2) histopathological report from a dedicated genitourinary pathologist of the Pathology Unit of the institution. These criteria ensured that the study focused on specific patients and lesions, providing a clear and targeted sample for analysis. By including only those who met these criteria, the researchers could ensure the reliability and validity of their findings, as well as the ethical considerations of the study. The exclusion criteria further refined the study population, eliminating potential confounding factors and ensuring the accuracy of the results.

Q: What is the role of expert radiologists in mpMRI-TRUS Fusion biopsy?

Expert radiologists play a crucial role in mpMRI-TRUS Fusion biopsy. They are highly qualified and experienced, with a minimum requirement of reading at least 250 mpMRI per year and performing at least 50 mpMRI-TRUS fusion biopsies annually. In the biopsy procedure, they utilize the mpMRI-TRUS Fusion image guide, a nondisposable biopsy gun, and a US platform with an end-fire TRUS probe. Their expertise ensures accurate and precise biopsy procedures, leading to reliable pathological examination results.

Q: How were ROIs segmented in mpMRI studies?

In mpMRI studies, ROIs were segmented by two experienced radiologists using ADC maps and T2w images. They manually contoured the prostate gland and each PI-RADS 3 lesion. After proper windowing, each radiologist, blinded to biopsy outcome, segmented the ROIs on all visible slices using ImageJ v1.53 software. This process ensured accurate and reliable segmentation for further analysis and predictive modeling.

Q: What is the two-stage procedure for radiomic feature generation?

The two-stage procedure for radiomic feature generation involves first computing 12 first-order features within each prostate ROI using a 9x9 pixel window. These features include mean, median, entropy, uniformity, standard deviation, median absolute deviation, interquartile range, coefficient of variation, skewness, kurtosis, and mean and median of the last decile of image pixel distribution. The second stage involves extracting feature values from PI-RADS 3 lesion ROIs and describing their global distributions through the same 12 first-order features. This yields 144 radiomic features (RFs) and 156 RFs when derived from ADC maps within lesion ROIs.

Q: What is the feature selection procedure for radiomic signatures?

The feature selection procedure for radiomic signatures consists of two main stages. First, after RF standardization, a preliminary subset of the most informative RFs arises from the least absolute shrinkage and selection operator (LASSO) regression, employing 10-fold cross-validation (CV) at the minimum CV error rule, and weighing each sample by its prior probability. Second, the most informative combination of four RFs is derived from the LASSO subset, by performing a preliminary discriminatory study. In this study, all possible combinations of 4 different RFs are assessed through support vector machines (SVMs). The discriminatory performance is ranked by the p-value of the Wilcoxon rank-sum test (p < 10^-3), corrected by Holm-Bonferroni, and in case of equivalent p-values, by the maximum Youden index (Y.I.) of the corresponding receiver operating characteristic (ROC) curve of each SVM-based diagnostic test, which is equal to specificity (SP) + sensitivity (SN) - 1. This procedure is repeated twice, using 1st- and 2nd-order polynomial SVM kernels, resulting in two radiomic signatures used for developing prostate cancer (PCa) diagnostic models.

Q: How was data augmentation used in the PCa diagnostic model training?

Data augmentation was employed to strengthen the statistical significance of the data subsets used for training, validation, and test phases. The initial dataset was augmented by 60%, resulting in an oversampled dataset (OD) comprising 248 samples, split between 114 PIRADS 3 benign lesions and 134 PIRADS 3 PCa. This augmentation process enhanced the dataset's robustness and improved the model's performance.

Q: What is the procedure for selecting the final PCa diagnostic model?

The procedure for selecting the final PCa diagnostic model consists of two main stages. First, among each run, one SVM classifier model with the highest Area Under the ROC (AUC) on the test fold is chosen, after discarding models prone to overfitting. This stage reduces the number of competing models to 100. Then, all 100-sample distributions of the SVM kernel parameters are considered, and the median value of each distribution is used to train the final optimized model. A separating hyperplane is defined for each SVM kernel, and its corresponding ROC curve is assessed. The SVM bias term is adjusted to meet the radiomic score at the Y.I. of the ROC. The PCa diagnostic models (linear and 2nd-order polynomial kernels) are tested on the holdout subset, and model performance is measured through SN, SP, I, PPV, and NPV. The performance is visually assessed through boxplots and waterfall plots. Finally, the radiomic score on the holdout test subset is converted into posterior probability using a binomial logit function.

Question

1. What is the role of mpMRI in detecting PCa in PI-RADS 3 lesions?

2. What is the aim of the study on mpMRI-based radiomic PCa diagnostic model?

3. How can radiomics improve the diagnosis of prostate cancer in PI-RADS 3 lesions?

4. What were the inclusion criteria for the study population?

Accepted Answer

The study investigates the role of mpMRI as a stand-alone tool for early and non-invasive detection of PCa in PI-RADS 3 lesions. It uses radiomic analysis of Apparent Diffusion Coefficient sequences. The methodology improves predictive performance, achieving a positive predictive value of 80%, with specificity = 76% and sensitivity = 78%. This suggests that mpMRI can be a valuable tool in identifying PI-RADS 3 lesions with a high probability of PCa.

Accepted Answer

The study aims to develop a multiparametric magnetic resonance imaging (mpMRI)-based radiomic prostate cancer (PCa) diagnostic model for PI-RADS 3 lesions. It enrolled 133 patients with 155 PI-RADS 3 lesions, 84 of which had PCa confirmation by fusion biopsy. Local radiomic features were generated from apparent diffusion coefficient maps, and the four most informative were selected using LASSO, the Wilcoxon rank-sum test (p < 0.001), and support vector machines (SVMs). The selected features were augmented and used to train an SVM classifier, externally validated on a holdout subset. Linear and second-order polynomial kernels were exploited, and their predictive performance compared.

Accepted Answer

Radiomics, which involves the use of machine learning and artificial intelligence strategies to analyze radiological images, can enrich the information retrieved by visual analysis. By generating quantitative measurements called radiomic features (RFs), radiomics can potentially diagnose benign and malignant lesions in PI-RADS 3 lesions. This approach can improve the predictive capability of the PI-RADS 3 score, allowing for more accurate diagnosis of prostate cancer. Previous studies have shown promising results in using radiomics for this purpose, and the aim of this study is to develop a machine learning model that predicts prostate cancer in a selected cohort of equivocal PI-RADS score 3 lesions, further improving the current state of the art through an automated pipeline of image processing and automatic feature generation.

Accepted Answer

The inclusion criteria for the study population were as follows: (1) MRI-TRUS fusion targeted biopsy (fusion-TB) only performed at the Radiology Unit; (2) histopathological report from a dedicated genitourinary pathologist of the Pathology Unit of the institution. These criteria ensured that the study focused on specific patients and lesions, providing a clear and targeted sample for analysis. By including only those who met these criteria, the researchers could ensure the reliability and validity of their findings, as well as the ethical considerations of the study. The exclusion criteria further refined the study population, eliminating potential confounding factors and ensuring the accuracy of the results.

Accepted Answer

Expert radiologists play a crucial role in mpMRI-TRUS Fusion biopsy. They are highly qualified and experienced, with a minimum requirement of reading at least 250 mpMRI per year and performing at least 50 mpMRI-TRUS fusion biopsies annually. In the biopsy procedure, they utilize the mpMRI-TRUS Fusion image guide, a nondisposable biopsy gun, and a US platform with an end-fire TRUS probe. Their expertise ensures accurate and precise biopsy procedures, leading to reliable pathological examination results.

Accepted Answer

In mpMRI studies, ROIs were segmented by two experienced radiologists using ADC maps and T2w images. They manually contoured the prostate gland and each PI-RADS 3 lesion. After proper windowing, each radiologist, blinded to biopsy outcome, segmented the ROIs on all visible slices using ImageJ v1.53 software. This process ensured accurate and reliable segmentation for further analysis and predictive modeling.

Accepted Answer

The two-stage procedure for radiomic feature generation involves first computing 12 first-order features within each prostate ROI using a 9x9 pixel window. These features include mean, median, entropy, uniformity, standard deviation, median absolute deviation, interquartile range, coefficient of variation, skewness, kurtosis, and mean and median of the last decile of image pixel distribution. The second stage involves extracting feature values from PI-RADS 3 lesion ROIs and describing their global distributions through the same 12 first-order features. This yields 144 radiomic features (RFs) and 156 RFs when derived from ADC maps within lesion ROIs.

Accepted Answer

The feature selection procedure for radiomic signatures consists of two main stages. First, after RF standardization, a preliminary subset of the most informative RFs arises from the least absolute shrinkage and selection operator (LASSO) regression, employing 10-fold cross-validation (CV) at the minimum CV error rule, and weighing each sample by its prior probability. Second, the most informative combination of four RFs is derived from the LASSO subset, by performing a preliminary discriminatory study. In this study, all possible combinations of 4 different RFs are assessed through support vector machines (SVMs). The discriminatory performance is ranked by the p-value of the Wilcoxon rank-sum test (p < 10^-3), corrected by Holm-Bonferroni, and in case of equivalent p-values, by the maximum Youden index (Y.I.) of the corresponding receiver operating characteristic (ROC) curve of each SVM-based diagnostic test, which is equal to specificity (SP) + sensitivity (SN) - 1. This procedure is repeated twice, using 1st- and 2nd-order polynomial SVM kernels, resulting in two radiomic signatures used for developing prostate cancer (PCa) diagnostic models.

Accepted Answer

Data augmentation was employed to strengthen the statistical significance of the data subsets used for training, validation, and test phases. The initial dataset was augmented by 60%, resulting in an oversampled dataset (OD) comprising 248 samples, split between 114 PIRADS 3 benign lesions and 134 PIRADS 3 PCa. This augmentation process enhanced the dataset's robustness and improved the model's performance.

Accepted Answer

The procedure for selecting the final PCa diagnostic model consists of two main stages. First, among each run, one SVM classifier model with the highest Area Under the ROC (AUC) on the test fold is chosen, after discarding models prone to overfitting. This stage reduces the number of competing models to 100. Then, all 100-sample distributions of the SVM kernel parameters are considered, and the median value of each distribution is used to train the final optimized model. A separating hyperplane is defined for each SVM kernel, and its corresponding ROC curve is assessed. The SVM bias term is adjusted to meet the radiomic score at the Y.I. of the ROC. The PCa diagnostic models (linear and 2nd-order polynomial kernels) are tested on the holdout subset, and model performance is measured through SN, SP, I, PPV, and NPV. The performance is visually assessed through boxplots and waterfall plots. Finally, the radiomic score on the holdout test subset is converted into posterior probability using a binomial logit function.

Accepted Answer

The most discriminant combination for PCa lesions consists of the standard deviation computed on the parametric map of the local median (M-s, RF55), the mean computed on the map of the mean value of the last decile (u 90th -u, RF61), the entropy of the map of the median value of the last decile (M 90th -e, RF83), and the s of the skewness map (S-s, RF91). These combinations were identified through discriminatory studies performed by linear SVM and polynomial SVM, with the best combination having a p-value of ~10^-11 and an I value of 0.57. The selected RFs are crucial for predicting PCa lesions and have shown promising results in terms of sensitivity and specificity.

Accepted Answer

Our study represents a proof-of-concept of the determinant contributions stemming from the ML analysis of mpMRI images for improving the diagnostic capability of radiological images, with early and noninvasive tools. Notably, the benefits of ML analysis have been proved even on the diagnosis of PCa in a selected cohort of PI-RADS 3 lesions, whose clinical management is still uncertain. The methodology adopted bounded the possibility of overoptimistic predictive performance, thus allowing us to be confident on the robustness of results, although this needs to be confirmed on a wider population. Actively, the present study also has some limitations. First of all, this study was a retrospective analysis of a relatively small group of patients from a single institution, with a single scanner, and before its wider clinical application, our predictive model needs to be prospectively validated on a larger scale, with patients from other medical units using different MRI scanners. Second, a separate analysis of PZ and TZ lesions was not performed due to too insufficient a sample size, and this will be pursued on a wider population. Finally, we used biopsy as the reference standard, which can be affected by undersampling errors; thus, a correlation with surgical specimens can be more appropriate, although from a methodological point of view, this implies, at most, a negative bias on the performance of the model. A future development, requiring a larger cohort, could be represented by the integration of radiomic analysis and clinical data, such as PSAD, in order to improve the accuracy of the model predictive of PCa and enforce a tailored approach in selecting patients with PI-RADS 3 lesions who need a biopsy.

An Apparent Diffusion Coefficient-Based Machine Learning Model Can Improve Prostate Cancer Detection in the Grey Area of the Prostate Imaging Reporting and Data System Category 3: A Single-Centre Experience

Chat with Paper

AI Agents for this Paper

Most frequently asked questions

1. What is the role of mpMRI in detecting PCa in PI-RADS 3 lesions?

2. What is the aim of the study on mpMRI-based radiomic PCa diagnostic model?

3. How can radiomics improve the diagnosis of prostate cancer in PI-RADS 3 lesions?

4. What were the inclusion criteria for the study population?

5. What is the role of expert radiologists in mpMRI-TRUS Fusion biopsy?

6. How were ROIs segmented in mpMRI studies?

7. What is the two-stage procedure for radiomic feature generation?

8. What is the feature selection procedure for radiomic signatures?

9. How was data augmentation used in the PCa diagnostic model training?

10. What is the procedure for selecting the final PCa diagnostic model?

11. What are the most discriminant combinations of RFs for PCa lesions?

12. How can radiomic analysis improve the diagnostic capability of mpMRI images for prostate cancer detection?

Citations

Research on texture images and radiomics in urology: a review of urological MR imaging applications.

Artificial Intelligence in Newborn Medicine

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

Radiomic Pipelines for Prostate Cancer in External Beam Radiation Therapy: A Review of Methods and Future Directions

Development and validation of a novel nomogram to avoid unnecessary biopsy in patients with PI-RADS category ≥ 4 lesions and PSA ≤ 20 ng/ml

References

The 2014 International Society of Urological Pathology (ISUP) Consensus Conference on Gleason Grading of Prostatic Carcinoma: Definition of Grading Patterns and Proposal for a New Grading System.

ESUR prostate MR guidelines 2012

Recent Global Patterns in Prostate Cancer Incidence and Mortality Rates.

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.

Head-to-head Comparison of Transrectal Ultrasound-guided Prostate Biopsy Versus Multiparametric Prostate Resonance Imaging with Subsequent Magnetic Resonance-guided Biopsy in Biopsy-naïve Men with Elevated Prostate-specific Antigen: A Large Prospective Multicenter Clinical Study.

Related Papers (5)

A comparison of feature selection methods for machine learning based automatic malarial cell recognition in wholeslide images

Combined SVM and PCA to Recognize the Brain Function from fMRI Images

An Ensemble of Deep Support Vector Machines for Image Categorization

Automated classification of brain mri using colorconverted k-means clustering segmentation and application of different kernel functions with multi-class svm

Research on the Segmentation of MRI Image Based on Multi-Classification Support Vector Machine