2023 ARRS ANNUAL MEETING - ABSTRACTS

RETURN TO ABSTRACT LISTING


E2359. Machine Learning Classification of Prostate Cancer with Non-Gaussian Diffusion Models as Novel Radiomic Features
Authors
  1. Zhihua Li; University of Illinois Chicago
  2. Guangyu Dan; University of Illinois Chicago
  3. Meet Patel; University of Illinois Chicago
  4. Xiaohong Zhou; University of Illinois Chicago
  5. Karen Xie; University of Illinois Chicago
Objective:
The purpose of this study is to examine the utility of radiomic features of the non-Gaussian fractional order calculus (FROC) diffusion model parameters from multi-parametric magnetic resonance imaging (mpMRI) in machine learning classification of prostate cancer (PCa) grade.

Materials and Methods:
A total of 140 patients who underwent targeted MRI-guided transrectal ultrasound (TRUS) fusion prostate biopsy within 6 months of MRI were reviewed retrospectively. Regions of interest (ROI) were placed on target lesions on MRI based on PI-RADS criteria and then correlated to histology results. The final tumor count is a total 140 peripheral zone PCas based on the International Society of Urological Pathology (ISUP) grade group: 20 of ISUP grade 1, 70 of grade 2 and 3, and 50 of grade = 4. Diffusion-Weighted Image (DWI) sequences of each lesion were fitted into the FROC model to generate parameter maps for feature extractions using the radiomic extension of 3D slicer. A random forest (RF) classifier model was applied to predict PCa grade and weight feature importance. Receiver Operating Characteristics (ROC) analysis was used to evaluate the performance of the RF models in classifying PCa grade.

Results:
For each FROC parameter (D, ß, µ), 93 radiomic features were extracted yielding a total of 279 features. RF models of FROC features achieved an accuracy of 0.857 for predicting ISUP grade 1 PCa, 0.714 for grade 2 and 3, and 0.959 for grade =4. In Receiver Operating Characteristics (ROC) analysis, the performance of RF classifier models achieved an area under curve (AUC) of 0.711 ± 0.011 for distinguishing ISUP grade 1; 0.718 ± 0.013 and 0.881 ± 0.009 for grades 2 and 3, and grade =4, respectively, after 10-fold cross validation and correcting for class weight imbalance.

Conclusion:
Non-Gaussian diffusion models such as FROC may provide additional viable features for machine learning classification of PCa grade. Additional optimization of an RF model with different FROC parameter map gradients and a greater sample size of grade 1 PCa is required to confirm its utility in practice.