AI, radiomics predict prostate cancer aggressiveness from MRIs

A new technique combining radiomics and machine learning can predict the aggressiveness of prostate cancer (PCa), sparing patients from invasive biopsy, according to a study published Sept. 5 in Clinical Radiology.

The method combines radiomic features from dynamic contrast-enhanced magnet (DCE)-MRIs with automatic machine learning approaches to classify how aggressive a lesion may be prior to invasive biopsy. B. Liu, with The Second Affiliated Hospital of Chongqing Medical University in China, and colleagues, believe their method may be included in the new version of PI-RADS.

“Assessing PCa invasiveness as early as possible is essential for disease management, treatment choice, and patient prognosis,” the authors wrote. “The present study investigated the clinical value of combining radiomics and automatic machine learning of original multiphase DCE-MRI images for discriminating PCas with low/intermediate or high invasive potential (GS ≤7 versus GS ≥8) before biopsy.”

The researchers included 40 patients who underwent a biopsy within four weeks of DCE-MRI exams. All patients had prostate cancer confirmed via biopsy. Lesions were segmented according to the time-signal-intensity curve and performed on the first and strongest phase of the enhancements on original DCE-MRIs. More than 1,000 radiomic features were taken from each lesion. This produced three datasets: Dataset-F, Dataset-S and Dataset-FS.

Five machine learning approaches were tested using cross-validation and evaluated by the area under the receiver operating characteristic curve (AUC).

Overall, 10 features from Dataset-FS produced the most accurate AUC of 0.93, and showed positive correlation with Gleason scoring. Dataset-F—which contained 40 regions of interest from the first enhancement phase—was “generally” better than the accuracy of Dataset-S, which contained 40 ROIs from the strongest enhancement phase.

“This method can facilitate the diagnostic ability to predict PCa invasiveness with an accuracy of 0.90 with no complications of infection/sepsis, haematuria, haematospermia, or rectal bleeding,” the researchers wrote. “It is a valuable means to help clinicians determine the appropriate treatment for their patients and has wide applicability in clinical practice.”

Among the several limitations in this study were the lack of separate training, test and validation sets and a small sample size. However, the team did address the validation issue by using fivefold cross-validation, they noted.

“As a new technology, this method is independent of conventional clinical assessment,” the group concluded. “It may be an effective complement to conventional MRI and included in the new version of PI-RADS in the future.”