Gaetan Dissez

MSc

behold.ai

Machine Learning Engineer

NLST (Learn more about this study)

NLST-799

Jun 14, 2021

Evaluation of a deep-learning algorithm to assist in the detection of features suggestive of lung cancer on chest x-rays.

At behold.ai, we have developed a deep-learning algorithm to detect features suggestive of lung cancer on chest radiographs. This algorithm has already been trained and evaluated with data coming from multiple sites and countries, demonstrating its efficacy and generalizability across different cohorts and characteristics.

However, several factors can restrain a thorough evaluation of deep-learning models for chest x-rays. In particular:
• The evaluation set is often limited to a small portion of the whole set of images available for the development of the algorithm.
• The evaluation of a deep learning algorithm for imaging often lacks more information about patients’ clinical outcomes and about the follow-up diagnoses that confirm or deny the presence of features suggestive of lung cancer at the chest x-ray level.

We believe that NLST datasets and chest x-rays can help us alleviate those restraining factors. We will evaluate and tune our models taking into account not only radiologists’ findings on chest radiographs but also follow-up exams and diagnoses. Indeed, features suggestive of lung cancer on chest x-rays can turn out to be confirmed cancers: those cases are particularly important to detect at the chest x-ray level. This evaluation will make our products safer and more efficient for patient care.

Our first goals are to evaluate our existing models on the NLST chest x-rays. The evaluation will consist of assessing:
• The performance of our models in detecting features suggestive of lung cancer on chest x-rays.
• The performance of our models in detecting other pathologies on chest x-rays.
• The ability of our models to localize nodules and masses on chest x-rays.

Then, using the final indications and diagnoses provided by biopsy or CT scans, we will try to assess if our algorithm could have helped detecting cancers in earlier stages on chest radiographs, reducing the number of cases missed by radiologists.