What kind of algorithm is contextflow built on?
We develop and train our own (Deep) Convolutional Neural Networks, which are designed in a way to create “maps” (embeddings) where semantically-similar patterns are very close together and different patterns are not. Thus, we can then identify visually-similar cases by looking at which patterns of existing cases are very close to the current selected one. For those similar cases, we can look at all the clinical and relevant information connected to them.
Can you retrain your algorithm solely on our data?
We don’t plan on retraining on local data at the moment. We simply enable a search of your local data.
How do you ensure the quality of your methods?
We have a streamlined data curation and annotation process where, together with our radiology experts, we validate findings in the image. Annotations are created to an extent where we can benchmark the quality of our methods. We use a mixture of unsupervised, semi-supervised and supervised approaches, minimizing the amount of annotated data necessary to provide high quality image retrieval performance.
How do you ensure your system doesn’t have bias?
Excellent question! 1) It’s true, you might not have all the data in the dataset…you might work with populations where one type of disease is not overly prevalent. For this, you need to be sure that the database covers most or all of the relevant patterns. Coal-mining example: the difference between lung diseases in the US and Europe is probably not that much but there are examples where, say, breast tissue in European females is more dense than in asian women and there you would need different data sets. So there’s no perfect answer to this except to take the direction from the doctors to know in which instances we need different datasets. 2) It could be that the dataset is biased based on how people are reporting. We do not blindly take the data from the hospital but we also have a data curation process, so there’s an in between step that we do. 3) Let’s say there’s a new disease description, so the reference content will be updated. We still have the outdated database but we link to the reference database, which is always state of the art. 4) We continually validate the algorithms to avoid overfitting the model.
Does your technology replace radiologists?
No, our technology is designed to support radiologists during the diagnostic process and does not perform automated diagnosis. We simply help radiologists with difficult cases and help them to tackle their high workloads. The radiologist is still in the driver’s seat.