How multiple -omics merge to predict Gleason grading

Sunday, November 26 | 9:50 a.m.-10:00 a.m. | S1-SSNMMI01-6 | Room E350

A new multiomics machine learning model may perform better for the prediction of Gleason grading than current methods used to determine the prognosis and management of prostate cancer, according to this scientific paper.

Attendees of this session will learn how radiomics data, genomics data, and pathomics data combined into a machine learning-based approach in a study that included a two-year follow-up of 146 patients. These were patients who underwent gallium-68 (Ga-68) prostate-specific membrane antigen (PSMA)-11 PET/MRI before radical prostatectomy between May 2014 and April 2020.

The research team led by Prof. Lukas Kenner and Alexander Haug, that included Jing Ning, who is affiliated to Christian Doppler Lab for Applied Metabolomics at the Medical University of Vienna, sought to improve the predictive performance of whole-mount Gleason grading in prostate cancer for a “smart and better stratification for radical prostatectomy” by deploying a high-throughput, machine-learning model to merge multiple “omics,” according to the RSNA abstract. For the machine-learning model, the team gathered this data using three strategies:

  1. Radiomics data acquisition involved delineated Ga-68 PSMA-11 PET/MR images from Hermes software with features computed using PyRadiomics.
  2. Genomics data acquisition involved isolating DNA from formalin-fixed paraffin-embedded (FFPE) samples from the radical prostatectomy. Ning and colleagues noted that whole-exome sequencing analysis was conducted, and functional pathway states were quantified using established in-silico tools. Clinical endpoints were also defined.
  3. Pathomics data acquisition involved constructing tissue microarrays from RP specimens, and immunohistochemical analysis was performed on 2-5-µm-thick sections from FFPE samples.

From these efforts, five machine learning-based models were established. The classification results were validated using 100-fold Monte Carlo cross-validation.

How well did the algorithm perform? You'll need to watch the presentation on Sunday to find out.

Page 1 of 2
Next Page