We at qure.ai, are transforming radiology, one patient at a time. We’re building AI & Deep Learning based expert medical imaging systems to capture abnormalities across multiple modalities – CT, X-Rays, MRI, Ultrasound etc.

Why AI

In case of an emergency, AI systems developed by Qure can easily and quickly diagnose the abnormalities, facilitating easy decision making for surgeons and patients likewise. The diagnosis helps in shortening the door-to-needle time for critical patients, increasing their odds for survival.

Also, expert systems at Qure helps in triaging of patients using algorithms to do basic screening. In cases of developing nation like India with acute shortage of radiologists in remote parts of the country, such screening systems can multiply the productivity of over-burdened radiologists.

Moreover, AI systems at Qure can quickly segment out multiple complicated lesions and quantify them volumetrically very accurately. The precise volume quantification helps in monitoring the progression of such lesions.

How to trust AI

All these computer vision based AI systems can diagnose certain abnormalities with precisions and recall comparable to human radiologists. However, a big gap lies between accurate predictions and good diagnosis – the ability to attribute. A good diagnosis not only detects abnormality but also explains why the radiologist thinks it’s abnormal. It becomes even more so important, for a clinician validating results of an automated system – as he needs to make sure that the algorithm takes the right decision by examining the right regions. For example, saying a patient has nodule, a clinical condition of presence of cancerous tissue in lung, doesn’t suffice alone if the algorithm can’t say where the nodule is or why there might be presence of a nodule.

Attributing decisions is the foremost problem for any deep learning based system. It becomes even more critical when it comes to medical imaging. Apart from the higher stakes involved in healthcare, the other reason attribution is critical for radiology is simply because they don’t intuitively make sense to us.

We can see an image of natural scenery and if a DL algorithm predicts it as containing bird objects – we can easily attest to the same. However, if a DL algorithm diagnoses a lung nodule in a chest CT scan, it can take some pondering even for trained eyes to detect the same. Also, as deep learning scientists ourselves without a thorough background in radiology, the opaque nature of classification nets makes debugging extremely testing.

Building trust

There are clearly 2 ways to build computer vision based systems we can trust:

  1. Build DL based systems to segment out relevant regions within the scan and then apply rule based system to diagnose abnormality
  2. Build classification systems where decisions can be visualized in form of a heatmap, indicating regions, in order of their importance, which contributed to the decision

The first one seems relatively simple, however there two major challenges involved. Firstly, building segmentation engines requires high quality detailed annotation data which is hard to obtain or generate. Secondly, such systems would be limited in terms of patterns that can be designed by radiologists, thus in effect limiting the capabilities of an innovative solution.

As regards to second point, we at Qure have been doing some cutting-edge research to build visualization systems to build AI systems we can trust.

The basic principle underlying such methods involve back-propagating the decision back to the input layer for classification nets. One of research abstracts, based on diagnosing abnormalities in Xrays and attributing the same in form of heatmaps, got accepted in the upcoming annual international conference organized by Radiological Society of North America (RSNA). We would soon be open sourcing some of these methods for others to adopt it as per their requirement.

Keep following us on our blog, Twitter & Github to keep yourself updated about the latest research we put out.

Editor's Note:
It is endearing to know one such innovative startup from India. As Rohit mentioned in the post, Qure.ai is aiming at open-sourcing some of its Research, kindly help us spread the word by commenting and sharing. And if you're an AI startup that is creating new pathways in your journey, please do say Hi!. Thank you for joining us in our journey TowardsAI.

Rohit Ghosh

Rohit is an 2014 IIT Bombay graduate, and is a Deep Learning scientist at Qure.ai ( http://qure.ai/ ), an AI startup in healthcare. His work is up there on: https://rohitghosh.github.io/