Methods of assessing lung disease in chest X-rays
The present system provides methods and systems of detecting lung abnormalities in chest x-ray images using at least two neural networks.
1. A diagnostic method, the method comprising: providing an image file of a chest x-ray from a patient to a machine learning system that has been trained on training data only available at a plurality of sources separated by time and/or geography and the training comprises connecting the machine learning system to the plurality of sources at different times and/or locations, wherein the machine learning system operates by resizing the image file of the chest x-ray into a first image that depicts the entire x-ray at a reduced resolution and placing a subsection of the file into a second image at an original resolution; analyzing the first and second images in parallel by respective first and second neural networks to output scores indicating a probability of a nodule, wherein the machine learning system has been trained to learn associations between features in chest x-rays and known pathology results with an area under the curve (AUC) of true positives over false positives for learned feature associations is between 0.7 and 0.83; and operating the machine learning system to detect lung nodules.
2. The method of claim 1 , wherein connecting the machine learning system to the plurality of sources provides a federated learning model by which patient data does not leave any of the plurality of sources.
3. The method of claim 2 , wherein the machine learning system comprises a plurality of distributed machine learning (ML) subsystems using a federated learning model to provide the machine learning system trained to detect abnormalities in chest x-rays.
4. The method of claim 1 , wherein the machine learning system is further trained on CT scans from the subjects to affirm or negate learned feature associations.
5. The method of claim 4 , wherein the CT scans are obtained as diverse CT scans from different instruments operating under diverse imaging conditions or parameters, the method further comprising reconstructing a standardized CT scan from each of the diverse CT scans.
6. The method of claim 5 , wherein the system includes a GAN trained to reconstruct standardized CT scans from the diverse CT scans and to augment the training data to make the neural networks generalizable across a plurality of image acquisition equipment and parameters.
7. The method of claim 1 , wherein the AUC≥0.74 is achieved by unsupervised learning without the aid of human once the learning is initiated.
8. The method of claim 1 , further comprising a human-in-loop training step that includes: displaying the chest x-ray to a clinician;
receiving at the machine learning system an annotation from the clinician that identifies a lung nodule in the chest x-ray;
comparing by the machine learning system the annotation to one of the scores indicating a probability of a nodule; and
using the comparison to improve learned feature associations.
9. The method of claim 1 , further comprising the machine learning system providing an alert that the system detected an abnormality in the chest x-ray from the subject and/or providing a region proposal including the abnormality in the chest x-ray.
10. The method of claim 9 , further comprising triaging the chest x-ray for immediate review by a human.
11. The method of claim 1 , further comprising characterizing an identified lung nodule with the machine learning system.
12. The method of claim 11 , wherein characterizing includes: classifying a nodule as a tumor, benign and/or malignant, and/or assessing or predicting nodule progression, volumetric sizing, nodule etiology, nodule histology, and/or response to treatment.