AI for lung cancer diagnosis
Neural network recognized lung cancer better than radiologists
But its effectiveness still needs to be checked
Evgenia Shcherbina, "The Attic"
Specialists from several US medical centers have trained a convolutional neural network to detect lung cancer on MRI images. In the tests, she showed herself either on a par with six radiologists, or better than them.
Lung cancer is the most common of oncological diseases. In 2018, out of 9.6 million deaths from them, 1.76 million were due to lung cancer. In the early stages, it is detected by screening, and then treatment can begin earlier, which increases the chances of the patient. To do this, doctors do an MRI of the chest, on which a specialist determines dangerous neoplasms in the lungs. However, the percentage of medical error in screening is still high – false positive or false negative results.
Therefore, scientists are looking for ways to improve the effectiveness of screening. One possible approach is to train a deep neural network to identify neoplasms in screening images and replace them in this way. In the course of their new work, specialists from several American medical centers took up this task. To do this, they trained a deep convolutional neural network on images from the NLST (National Lung Screening Trial, National Lung Cancer Screening) database. These are 42,290 MRI images taken from 14,851 patients. In 578 of them, cancer was accurately detected by biopsy (taking tissue samples).
The model works as follows: lung tomograms and, if available, earlier images (for example, from the previous screening) are loaded into it. It analyzes the suspicious areas in these images and provides a general malignancy prognosis for a specific case, an assessment of the risk of cancer and the localization of the suspected cancerous area.
The model was tested on images of 6719 patients (89 of them were accurately diagnosed with cancer) and compared with the results of six experienced radiologists who were provided with a set of images of 597 patients (83 cases of cancer). In the case when the model and radiologists had only one set of images, she surpassed all six specialists, reducing the number of false positive results by 11% and false negative by 6%. The overall accuracy of cancer detection was 94.4%.
If specialists and neural networks had previous pictures (and people also had a medical history and other data about the patient), then the neural network was as accurate as people. The result was confirmed by an independent clinical examination in 1139 cases.
The authors warn that these results still need to be tested on large samples of patients, but hope that the model will be able to help specialists diagnose cancer.
Portal "Eternal youth" http://vechnayamolodost.ru