OUR ARTIFICIAL INTELLIGENCE LED APPROACH

Optellum LCP (Lung Cancer Prediction)* is a digital biomarker based on Machine Learning that predicts malignancy of an Indeterminate Lung Nodule from a standard CT scan.

  • AI-based digital biomarker – computed from CT images only.
  • Trained on more than 100,000+ datasets with known diagnostic ground truth.
  • Produces a malignancy score between 1 (benign) and 10 (malignant) for a nodule of interest.

HOW IT WORKS

When a user marks a nodule in the Virtual Nodule Clinic, the software automatically sends the relevant parts of the CT scan to the AI (Artificial Intelligence), which tests the nodule for malignancy and provides an LCP-CNN (Lung Cancer Prediction Convolutional Neural Network) score. For users, this AI score is conveniently grouped into 10 categories, corresponding to the least risky 10% of the population, the next most risky 10% and so on (at a usual lung nodule clinic cancer prevalence), and this is called the Optellum LCP Score. Thus a nodule scoring 4 has an exceedingly high chance of being benign, whereas a nodule scoring 9 is much more likely to be a cancer. This is shown in the bar chart below-right, which also appears for a user when they see a score.

Mark the nodule on the CT in Virtual Nodule Clinic

LCP-CNN

AI Lung Cancer Prediction

OPTELLUM
LCP SCORE
9

Convert to simple value to assist human reader.
Optellum score of 9 corresponds to 84% risk of cancer

How does the computer produce the malignancy score?

A. A stack of 2D images at different heights around the nodule is fed into a convolutional neural network. The layers of the network (depicted as boxes representing arrays of numerical weights) convert this stack to a single nodule descriptor (long vector of numbers), and then a classifier takes these numbers and outputs a single malignancy score that encapsulates exactly how malignant the network thinks this stack of input images is.

B. A network is trained using nodule data with known malignancy status. Nodule image stacks are fed through the network, and all the places where the network misclassifies a nodule are identified. Those errors are then fed back through the network, updating its numerical weights in order to make fewer mistakes on the next pass. After thousands upon thousands of passes, the network develops the ability to tell cancer from benign disease more and more accurately.

C. In testing mode, we don’t know whether a nodule is cancer or not, but we have shown that by this point, the network training has made it very accurate. Thus, new data can be fed through, and instead of looking for misclassifications, we use the output from the network as a prediction of how malignant the nodule is. This LCP-CNN score ranges from 0 to 100, and is used by the Virtual Nodule Clinic to give a nodule category ‘Optellum LCP score’ from 1-10

Source: p.243 Assessing the Accuracy of a Deep Learning Method to Risk Stratify Indeterminate Pulmonary Nodules
Journal: American Journal of Respiratory and Critical Care Medicine (also known as ATS Blue Journal)

*Optellum LCP FDA 510(K) pending.