Machine learning model may influence individualized care of Crohn’s disease in children
Click Here to Manage Email Alerts
A machine learning image model describing Crohn’s disease phenotypes may impact individualized care for children with Crohn’s, according to a presenter at the Crohn’s and Colitis Congress.
“Our model predicts [Crohn’s] B1 vs. B2 vs. controls with histological features greater than 70% accuracy,” Sana Syed, MD, from the University of Virginia School of Medicine, said during her presentation. “The model decision making visualized via histological features showed connective tissue was possibly the distinctive feature for predicting B2.”
Syed and colleagues obtained baseline hematoxylin and eosin stained ileal biopsy slides from the Cincinnati Children’s Hospital Medical Center’s RISK validation sub cohort. The study also comprised 10 ileal histological controls. Researchers included B1 (non-stricturing, non-penetrating; n = 97) and B2 (stricturing; n = 71) data.
Biopsy slides were digitized at the University of Virginia and a ResNet101 CNN model was trained. For training, the high-resolution images were patched into 1000 x 1000 pixels with 50% overlap and resized to 256 x 256 pixels. The model’s decision-making process was visualized with Gradient Weighted Activating Mappings.
Investigators trained the model for CD compared with controls, which achieved 97% accuracy in the detection of controls. They further trained the model for classifying phenotypes of CD. The model demonstrated a higher accuracy in detecting B2; however, overlaps with other phenotypes were detected, Syed reported.
“We further want to train our models and validate them with larger sample sizes and more Crohn’s phenotypes,” she said. “Then, look at the B2 and B3 overlap.”