Canadian researchers developed IPRO, a deep learning model for lung cancer care, enhancing prognosis by analyzing pre-treatment CT scans without manual input.

Deep learning is making significant strides in the field of medical imaging, particularly in lung cancer care. A Canadian research team has developed a fully automated imaging-based prognostication technique (IPRO) that utilizes deep learning to predict mortality risks for patients with stage I-IV lung cancer based on pre-treatment CT scans. This approach aims to standardize and enhance prognostic assessments while eliminating the subjectivity and variability associated with manual lesion analysis. In a retrospective study, IPRO demonstrated comparable prognostic power to traditional TNM staging and, when combined, provided superior predictions of patient outcomes.

The IPRO model stands out from other deep learning models by not being trained to replicate specific radiological tasks but instead by predicting mortality risks without manual annotations. Dr. Felipe Soares Torres, the assistant professor of radiology at the University of Toronto, emphasizes that IPRO is designed to assist radiologists rather than replace them, by providing additional, standardized prognostic information to inform personalized treatment decisions. The model’s accuracy and attention are measured by heatmaps and its associations with known prognostic factors, paving the way for its integration into clinical decision-making.

The potential clinical value of IPRO is recognized by oncology professionals who see it as a tool to refine treatment strategies and improve clinical trial design. Its comprehensive patient stratification could lead to better baseline randomization and more accurate associations with overall survival. Future research includes incorporating IPRO into decision-making algorithms and clinical pathways for evidence-based care in non-small cell lung cancer (NSCLC). The team is also exploring how to combine IPRO with other biomarkers and clinical variables to further refine its prognostic accuracy, with ongoing research on integrating information from pulmonary function tests into the model.

Read Full Article