Optellum Science at the IASLC World Conference on Lung Cancer

This week Optellum has been highlighted in three presentations at the IASLC World Conference on Lung Cancer in Toronto, Canada.

On Tuesday, Dr. Peter van Ooijen, from University Medical Centre Groningen presented the results of the EIT Health LUCINDA project in a talk entitled “Lung Cancer Prediction Using Deep Learning Software: Validation on Independent Multi-Centre Data”. This collaboration between University Centre Groningen, Heidelberg University Hospital & The Thoraxklinik, Heidelberg, Oxford University Hospitals and Optellum, has validated for the first time a Deep Learning based lung cancer prediction model (Optellum LCP) on independent real-world datasets collected from three different centres. The results showed that overall classification remains very high (AUC=0.92 95%CI = 0.89-0.93) despite the very different patient populations. Rule-out performance was also excellent with 25.1% of benign patients being correctly ruled out from one CT scan with a Negative Predictive Value of 99.5%.

A second abstract, entitled “AI Based Malignancy Prediction of Indeterminate Pulmonary Nodules: Robustness to CT Contrast Media” presented as a poster-presentation with Oxford University Hospitals collaborators, examined the robustness of the Optellum Lung Cancer Prediction algorithm to the presence of contrast media in the CT study. Pulmonary nodules detected incidentally may present in a very wide variety of imaging studies; examples include CT angiography, contrast-enhanced diagnostic CT and PET/CT. Therefore it is important that any rule-out test is robust to the presence of contrast media in the image. The study found the Optellum LCP is robust to many contrast enhanced scans except for those with very high levels of enhancement.

The third abstract, entitled “Automatic Nodule Size Measurements Can Improve Prediction Accuracy Within a Brock Risk Model”, with Oxford University Hospitals, looked at the effect of using an automatic nodule volumetric segmentation algorithm of the classification performance of a logistic regression multi-parametric nodule malignancy risk model. Such models typically rely on the patient and nodule specific parameters to predict nodule malignancy, where nodule diameter is one of the key input parameters. The study found that utilizing automatic measurements of nodule diameter improved the classification performance from 86.5% to 88.5% on a dataset of 5373 nodules, an improvement of just over 2 AUC points.