Combining AI and Radiomics to Predict Prostate Cancer Aggressiveness
The combination of radiomics and an artificial intelligence (AI) algorithm can accurately and automatically predict the aggressiveness of prostate cancer (PCa) prior to biopsy, according to research by Liu et al,1 published online September 5, 2019 in Clinical Radiology. Radiomics technology uses data characterization algorithms to examine a vast number of quantitative features from medical images. Machine learning is a form of AI that employs statistics to allow computer systems to learn from data and automatically improve performance. Rates of prostate cancer are increasing. Although most tumors grow slowly, some are more aggressive, making accurate diagnosis crucial for managing disease and improving outcomes.2,3
Through this study, the authors aimed to establish the clinical value of radiomics plus AI learning of original multiphase dynamic contrast-enhanced MRI (DCE-MRI) images to accurately and noninvasively stratify PCa according to invasive potential, prior to biopsy.
Methods and Findings
From January 1, 2016 until May 31, 2018, 50 eligible patients with PCa and a Prostate Imaging-Reporting and Data System (PI-RADS, version 2) score of 4 or 5 were identified. All underwent MRI and biopsy within 4 weeks. Exclusion criteria were poor MRI image quality, cancer location requiring a urethral catheter, multiple lesions with different Gleason scores, and previous PCa treatment. Ultimately, 40 lesions from 40 patients were included in the study.
Tumor-lesion time-signal-intensity curves were calculated using a General Electric Medical Systems advanced workstation. Regions of interest (ROIs) were identified slice by slice during the first phase of enhancement and during the strongest phase on the original DCE-MRI images. Ultimately, 40 ROIs from the first phase were grouped into Dataset-F, 40 ROIs from the strongest phase were grouped into Dataset-S, and 80 ROIs from both phases were grouped into Dataset-FS.
Overall, 1029 quantitative radiomics features were extracted from each ROI. The feature dimensions were reduced using the variance threshold method, select k-best method, and least absolute shrinkage and selection operator (LASSO) algorithm. This procedure identified the optimal subset of features from each dataset to utilize for the machine-learning model: eight from Dataset-F, four from Dataset-S, and 16 from Dataset-FS. Five machine-learning methods were employed independently for cross-validation. Correlation analysis compared the features of the machine learning model that achieved the best classification performance. In addition, another comparison was related to the Gleason score of each PCa lesion.
According to logistic regression analysis, Dataset-FS had the highest predictive value of the three datasets. Ten of the 16 features in Dataset-FS showed significant correlation with Gleason scores: F-least axis shape, S-least axis shape, F-total energy first-order statistics, F total energy logarithm, S-large area emphasis in GLSZM-wavelet-LLH, S-large area high grey-level emphasis in grey level size zone (GLSZM) texture features, F-zone entropy in GLSZM-wavelet-HHL, F-long run emphasis in (grey-level run length matrix) GLRLM-wavelet-HHH, F-run length non-uniformity in glrlm-exponential, and S-long run emphasis in GLRLM-wavelet-HHH. With respect to the two other datasets, F performed better than S. This may be due to the fact that the initial phase of tumor enhancement has more diverse information than the strongest phase of enhancement.
The DCE-MRI imaging technique offers many scanning phases, but consensus is lacking as to which is best for predicting pathological response. This study utilized the first and strongest phases of tumor enhancement, which the investigators believe may have the greatest association with potential tissue information. Other research suggests that the strongest phase can be indicative of tumor invasiveness.4
The authors acknowledged that their study has several limitations, including the relatively small sample size and the failure to investigate peripheral zone and transitional zone neoplasms separately. Most Gleason scores were derived from transperineal ultrasound-guided biopsy specimens, which may differ from samples obtained during radical prostatectomy. The researchers called for future work to confirm their results, using larger sample sizes, and to compare their findings with those of other imaging techniques.
- Liu B, Cheng J, Guo DJ, et al. Prediction of prostate cancer aggressiveness with a combination of radiomics and machine learning-based analysis of dynamic contrast-enhanced MRI. Clin Radiol. 2019 Sep 5. pii: S0009-9260(19)30355-1. Doi: 10.1016/j.crad.2019.07.011. [Epub ahead of print]. Accessed October 5, 2019.
- McGuire S. World Cancer Report 2014. Geneva, Switzerland: World Health Organization, International Agency for Research on Cancer, WHO Press, 2015. Adv Nutr. 2016;7(2):418–419. Accessed October 6, 2019.
- Heidenreich A, Bastian PJ, Bellmunt J, et al. EAU guidelines on prostate cancer. part 1: screening, diagnosis, and local treatment with curative intent-update 2013. Eur Urol. 2014;65(1):124–137. Accessed October 6, 2019.
- Teruel JR, Heldahl MG, Goa PE, et al. Dynamic contrast-enhanced MRI texture analysis for pretreatment prediction of clinical and pathological response to neoadjuvant chemotherapy in patients with locally advanced breast cancer. NMRI Biomed. 2014;27(8):887–896. Accessed October 6, 2019