The field of medicine faces a global shortage of radiologists, combined with increasing workloads and complex diagnoses, leading to delays, missed findings and huge overtime expenses in healthcare.
Contextflow brings together experts in medical imaging and artificial intelligence from the Medical University of Vienna (MUW) and the Technical University of Vienna to help radiologists prioritize and diagnose difficult cases faster and more accurately using Julia.
A radiologist will upload a scan and mark a region of interest. Contextflow’s three-dimensional image-based search engine searches a database of thousands of images to find similar cases based on visual disease pattern detection. Each result is a full volume, so radiologists can scroll through or change the contrast or brightness. Radiologists can even restrict results based on age, gender or pathological findings in the report via text search.
Contextflow uses Knet.jl, a Julia deep learning package, to build the models that enable Contextflow’s three-dimensional image-based search engine to identify reference cases for physicians.
Contextflow relies on complex data processing pipelines to deliver results to customers immediately with tight computing resources.
Clear and concise language
Malleable through metaprogramming
Great low-level deep learning libraries (Knet.jl)
Easy to achieve abstract (data flow graphs) and low-level tasks (deep learning engineering, input/output, control over memory layout)
Radiologists often spend up to 20 minutes searching for information to help them with a diagnosis. With our tool, the search time is cut down to ~2 seconds,
Donner explains.
We would not be where we are now without Julia. It is a joy to use Julia to express one’s ideas.
Furthermore,
Julia allows us to go from initial code to entire systems without the need to reimplement anything in a second language. We use Julia’s metaprogramming capabilities and deep learning frameworks that are simply not available in other languages.