ES01.03 Deep Machine Learning for Screening LDCT

Abstract

The first computer algorithms to automatically detect pulmonary nodules in CT scans, based on classical machine learning approaches, were developed almost two decades ago. These systems appeared in commercially available computer-aided detection packages. However, a recent study concluded that such older software systems fail to flag a substantial number of cancerous lesions and have a fairly high false positive rate. Recently, algorithms based on deep learning, in particular, convolutional neural networks, have been developed that report high sensitivity with low false positive rates. Similar deep learning algorithms have been successful in classifying nodules as solid, subsolid or part-solid with accuracy comparable to radiologists, and in estimating the probability of malignancy of nodules. The 2017 Kaggle Data Science Bowl combined these tasks into a single challenge where 2000 teams developed methods to predict, on the basis of a single screening CT scan, whether a patient would be diagnosed with lung cancer within one year of the date of the scan. The 10 best performing solutions are now available under an open source license and form the basis of commercial solutions that show, in recent validation studies, a performance comparable to radiologists. Thorough validation studies are now needed to investigate if the good performance of these deep learning systems can be replicated, independent of CT parameters, and how such systems can be implemented in a lung cancer screening setting. Possibilities include the use of AI software as a second reader, as a concurrent reader, or even a stand-alone reader for a fraction of the cases, when widespread implementation of screening will put a too large burden on scarce radiological resources. In this lecture, I will review the currently available computer solutions and discuss their validation and integration into CT lung screening workflows. artificial intelligence, deep learning, chest ct