Machine learning (ML) has been beneficial in medical imaging applications, and it brings an exciting opportunity to enhance the availability, latency, accuracy, and consistency of chest X-ray image interpretation. The plethora of algorithms that have already been developed to detect specific conditions, such as lung cancer, tuberculosis, and pneumothorax, are a great example. However, the application of these algorithms may be limited in a general clinical setting. A wide variety of abnormalities could surface here, making diagnosis more complicated than it initially seems.
For example, a pneumothorax detector is not expected to highlight nodules suggestive of cancer, and a tuberculosis detector may not identify findings specific to pneumonia. An initial triaging step is determining whether CXRs contain concerning abnormalities. Therefore developing a general-purpose algorithm that identifies X-rays having any sort of abnormality could significantly facilitate the workflow. But developing a classifier is challenging due to the variety of abnormal findings that present on CXRs.
Google recently published this new research, “Deep Learning for Distinguishing Normal versus Abnormal Chest Radiographs and Generalization to Two Unseen Diseases Tuberculosis and COVID-19”, published in Scientific Reports. Google’s new deep learning model can distinguish between normal and abnormal chest X-rays from the deidentified data set. This model performs well for general thoracic abnormalities and tuberculosis as it is universal to unseen cases such as COVID-19.
The research team used a deep learning system based on the EfficientNet-B7 architecture, pre-trained with ImageNet. They then applied this model to 200,000 de-identified CXRs from Apollo Hospital in India and Zhang Qu’s training data set of chest conditions for identification purposes.
The researcher team used various data sets to test the proposed deep learning system and they found that it can accurately distinguish common chest abnormalities. The researchers found that the system’s ability to detect unexperienced diseases is still very high, especially in tuberculosis, but it was slightly reduced in identifying COVID-19. The research team explained this by saying many cases were marked as abnormal by the system despite them not belonging to specific diseases, however, they could be distinguished as abnormalities.
Google’s new system has two significant benefits. In addition to accelerating the sorting of patients by the situation, it can also improve the effectiveness of other chest X-ray image recognition models. To understand the help of the model in clinical workflows, researchers use it to prioritize simulated cases. Abnormal cases can be prioritized and with quick transfer results to specialized radiologists for treatment; patients who need urgent treatments are classified quickly.
Google’s proposed deep learning system can also be used as a pre-training model. In limited data cases, other machine learning algorithms can be improved. Recently, Google used normal and abnormal detectors together with tuberculosis detectors to provide services for the early diagnosis of tuberculosis in areas lacking medical resources.
NIH Chest X-ray dataset with Labels: https://cloud.google.com/healthcare/docs/resources/public-datasets/nih-chest#additional_labels