A tool by Joseph Paul Cohen PhD and his team can read your chest X-ray and provide a “NOT FOR MEDICAL USE” second opinion. In the current version, it rates the likelihood of 18 different conditions based on its extensive training set database. In my case, it was very certain that I didn’t have any of the 18 conditions, but it indicated an increased possibility for three of them: a nodule, a fracture, and emphysema. I was still in the healthy range, but these were of increased possibility.
You do first need an X-ray image (frontal) of your chest. I took a screenshot of mine while reading it from Horos, where I also view my MRI images. Then, having an image of my chest X-ray, I download the software and used it without uploading my X-ray image anywhere; everything stays local on your computer, which is super awesome. It did only work for me with front images, not with side images or images and it did not work with a frontal chest image that was too long and not wide enough.
This tool will eventually be available for medical use, so keep an eye on Chester, the AI Chest X-Ray Radiology Assistant.
https://mlmed.org/tools/xray/
From the Site:
“In order to bridge the gap between AI researchers and medical professionals, we developed a very accessible free prototype system that can be used by medical professionals to understand the reality of deep learning tools for chest X-ray diagnostics. The system is designed to provide a second opinion, allowing users to process an image to confirm or aid in their diagnosis. The tool predicts 18 different radiological findings based on data from the seven largest public datasets. What makes this tool unique is that the web version runs entirely locally, and no data is sent off the device. This allows the tool to scale to millions of users for free. The tool is available as a webpage that works on computers and mobile phones, and with our new version 3 release, we provide native Windows and Mac versions.” – Read More [4]
From a FastCompany article:
“The Chester AI radiology assistant was developed in work led by Joseph Paul Cohen, a postdoctoral fellow at Mila (the Quebec AI institute) and the University of Montreal. He used an NIH dataset of chest X-rays and diseases to train software to spot diseases in these scans. Though he is not a clinical doctor, Cohen is focused on the intersection of health and deep learning. Previously, Cohen created an app called BlindTool, which used machine learning to train a phone’s camera to serve as the eyes for someone with vision impairment. The “This tool is for a second diagnosis. So far our interaction with doctors has been that it is useful if they are in a hurry (like in an ER) and want to have someone run this image to confirm what they think or to help them not miss anything,” says Cohen. “For radiologists in training, this will help them to form a consistent understanding no matter who their teacher is.” – Read More [3]
Eighteen Conditions Checked, What Each Is
The eighteen conditions which are scored in the version I tested, 3.2, were specificially:
- Atelectasis: The collapse of part or all of a lung, often due to blockage of air passages or pressure from outside the lung.
- Consolidation: The process where lung tissue becomes firm and solid due to the accumulation of fluid, cells, or other substances, often seen in pneumonia.
- Infiltration: The presence of abnormal substances or cells within the lung tissue, which can indicate various diseases, including infections or malignancies.
- Pneumothorax: The presence of air in the pleural space, leading to partial or complete collapse of the lung.
- Edema: The accumulation of excess fluid in the lungs, which can cause difficulty in breathing and is often associated with heart failure.
- Emphysema: A chronic lung condition characterized by the destruction of the alveoli (air sacs) in the lungs, leading to reduced airflow and breathing difficulties.
- Fibrosis: The thickening and scarring of lung tissue, which can result from various lung diseases and can impair lung function.
- Effusion: The accumulation of fluid in the pleural space surrounding the lungs, which can restrict lung expansion and cause breathing difficulties.
- Pneumonia: An infection that inflames the air sacs in one or both lungs, which may fill with fluid or pus, causing cough, fever, and difficulty breathing.
- Pleural Thickening: The abnormal thickening of the pleura (the lining around the lungs), which can indicate chronic inflammation or malignancy.
- Cardiomegaly: An enlargement of the heart, which can result from various conditions, including high blood pressure and heart disease.
- Nodule: A small, rounded mass of tissue in the lung that can be benign or malignant and is often detected on imaging studies.
- Mass: A larger abnormal growth in the lung that may indicate the presence of cancer or other serious conditions.
- Hernia: A condition where an organ or tissue protrudes through a weak spot in the surrounding muscle or connective tissue, which can occur in the diaphragm affecting the lungs.
- Lung Lesion: Any abnormal area of tissue in the lung, which can be due to infection, inflammation, or tumors.
- Fracture: A break in the rib or other bones in the chest area, which can cause pain and difficulty breathing.
- Lung Opacity: An area in the lung that appears denser than normal on imaging, indicating potential abnormalities such as fluid, infection, or tumors.
- Enlarged Cardiomediastinum: An increase in the size of the mediastinum (the central compartment of the thoracic cavity), which can indicate various conditions, including heart enlargement or tumors.
If one of these is rated as being high in the risk category, in other words if the tool is not certain that the image given matches the healthy images in its training set for a given diagnosis, a button appears to click so you can learn more. This button highlights the areas (heat map / or points) on your image which triggered the diagnosis.
To accomplish it’s A.I. magic, the tool used “DenseNet-121 (Input 224×224) xrv-all-45rot15trans15scale (TorchXRayVision)” as it’s Prediction Model[1] and “Autoencoder (Input 64×64) ae-chest-savedmodel-64-512 ” to determine if the image given is outside of the training data’s ability to give an answer. (Out of Distribtion Model)[2]. Training data sources include “NIH, PadChest, RSNA Pnuemonia, CheXpert, MIMIC-CXR”Best of all, the tool respects user privacy and no data from the user is sent off of the device.
If you are interested in how it works, here is more detail, including many things the developer had to account for to start to get accurate results.
While Chester is a great and amaizing tool, it does not measure what I set out to check for in my chest X-ray: osteoarthritis. I’m still looking for that one.
Read More
[1] https://arxiv.org/abs/2002.02497
[2] https://arxiv.org/abs/1901.11210
[3] https://www.fastcompany.com/90326445/this-free-ai-reads-x-rays-as-well-as-doctors
[4] https://mlmed.org/tools/xray/
[5] https://www.youtube.com/watch?v=oPoz63ACs9A