Google's New AI for Categorizing Chest X-Rays

2019-12-04

Analyzing chest X-ray images using machine learning algorithms may sound easy, but in reality, it is a challenging task. This is because the clinical data required to train these algorithms is mostly obtained through rule-based natural language processing or manual annotation, both of which often have potential inconsistencies and errors. Additionally, collecting a dataset that represents a diverse spectrum of cases and establishing clinically meaningful labels using only images is a challenging task. To make progress in the classification of X-ray images, Google researchers have designed an artificial intelligence model to identify four findings in human chest X-rays: pneumothorax (collapsed lung), nodules, masses, and lung opacities. In a paper published in the journal "Nature," the Google research team stated that their series of models demonstrated performance at the level of radiologists in an independent evaluation. A few months before this research was published, Google AI and scientists from Northwestern University created a model that outperformed radiologists with an average of 8 years of experience in detecting lung cancer through screening tests. About a year ago, New York University used Google's Inception v3 machine learning model to detect lung cancer. These two past facts illustrate the preparation and effort behind Google's success this time. The new artificial intelligence model utilizes over 600,000 images from two unverified datasets. The first dataset was developed in collaboration with Apollo Hospitals and consists of X-rays collected from multiple locations over the years. As for the second corpus, it is a publicly available set of ChestX-ray14 images released by the National Institutes of Health, which has been used as a resource for artificial intelligence but has accuracy deficiencies. The researchers developed a text-based system that extracts labels from radiology reports associated with each X-ray image and applied this system to provide labels for over 560,000 images from the Apollo Hospitals dataset. To reduce errors introduced by text-based label extraction and provide relevant labels for multiple ChestX-ray14 images, they recruited radiologists to review 37,000 images from both datasets. Three radiologists reviewed all the final tuning and test set images and resolved discrepancies through online discussions. The final result is that although the model achieved expert-level accuracy overall, its performance varies when faced with different databases. For example, the sensitivity for CHEST-Ray14 images is approximately 79%, but for the same images in other datasets, it is only 52%. The research team hopes to establish a better method by using the ChestX-ray14 dataset as a reference label database. Although the stability of this new model is still debatable, it can already be applied to specialized medical activities, marking another small step in the field of AI healthcare.

RECENT AI NEWS

RECENT AI TOOLS