NILG.AI

In case you missed it. You may find below our webinar on the Symposium on Bioengineering 2020 at FEUP. Also, we reply to some answers that we didn’t have time to answer during the talk.

Abstract

Despite being easily preventable, cervical cancer remains a leading cause of death among women — the main reason behind this being the limited access to high-quality screening programs by patients in remote locations. Artificial Intelligence and Mobile Computing are changing the way we give care to patients, enabling vulnerable populations to have access to such services. In this talk, we will discuss how we are using Deep Learning to fight cervical cancer together with MobileODT, an Israel-based company in the endeavor of equipping every healthcare provider, everywhere, with the most advanced tools to save as many lives, as quickly as possible. We will discuss both the technical and social challenges in this journey, from the data collection process to the proper validation of the technology.

Q&A Session

1. How was the process of introducing this device to the countries you showed? Was it easy to approach the governments and how accepting were they?
   • I was not involved in this part. However, I can tell you the process was conducted one by one, with hard work and perseverance. MobileODT has teams devoted to this and only this topic. They have regional teams handling local specificities as well as regulatory teams. For further information, you may address MobileODT and take a look at their blog.
2. In your perspective, what was the make or break factor in this project’s success?
   • The break factor in these applications is always non-technical. It was the people behind it, the doctors and nurses in the field, MobileODT as a whole, and the research community. From a technical point of view, the EVA colposcope was crucial as a mobile device that can be brought to remote locations. 
3. What’s your position in extending this artificial Intelligence in the treatment of other disorders? Meaning, do you believe in creating a super AI diagnostic assistant?
   • I don’t believe current advances in Deep Learning will lead us to that super AI. I suggest reading “The Book of Why” by Judea Pearl and “AI Compatible” by Stuart Russell. However, deep learning reduced the entry cost for building models that fill local gaps in the decision process. So, I believe that, in the short term, we may have an ecosystem of CAD systems that interact with each other to have a more holistic view of the patient. We have multiple challenges to address that, but I think it’s more attainable than a single super AI healthcare agent. If you want to read our vision in the area, go read our previous post on AI applications in healthcare.
4. What are the major constraints of applying these frameworks to other types of cancer or other medical images?
   • Before deep learning, the major constraints were the 1000 decisions you made when building the system. Now, the three main constraints are:
     • Data: Do you have enough data? is this data biased somehow? do you have labels that are strongly correlated with your KPIs?
     • Business Integration: Where in the decision process/workflow are you deploying your models?
     • Domain Knowledge: What domain-specific knowledge differentiates each area and how can we embed this information so the models don’t need to learn that from scratch.
5. CNNs have been used with very good results for image/pixel classification in various fields, but with different ranges of attainable performance in the different medical applications. For this one, histology imaging for cancer detection, what is the minimal error (aka Bayes error, or best possible performance) you expect to possibly obtain? Thank you!
   • We don’t even know what’s the actual prevalence of cervical cancer worldwide. The world is under-measured in many aspects. So, we are not sure what would be the minimal error to expect, I hope it is low enough to ensure a rookie practitioner on a low-resource location can provide the same level of care than a rockstar colposcopist.
6. Have you thought about incorporating other types of data into the classification model, such as text from medical records and personal data form patients?
   • Yes, actually we have done that in our previous research. We have developed multimodal models that, besides images, include other medical records from the patient, risk factors, among others. Also, we have done this not only in the input but also on the model output. For instance, incorporating additional tasks that may be complementary to support the decision.
7. You mentioned the dataset that you were using. Is that dataset large enough to assure a good quality diagnosis?
   • So far, it’s the largest dataset in the field. Moreover, it’s probably larger than the number of patients any human doctor can see during her working life. Good quality goes hand in hand with reproducibility and AI can provide that. In this sense, we conducted some experiments to validate the performance of the models on several adversarial conditions.
8. Will the data be manageable/understandable for the doctors in these countries? And is there any sort of treatment there or will the patients when diagnosed be transported to other countries?
   • • On the technical side, I always suggest that AI systems should go beyond predictions (i.e., probabilities, classes). We focus on transforming predictions into decisions and decisions into actions. While a score might be an abstract concept for a non-technical audience, supporting evidence is an actionable tool. Thus, we focus on building explainable AI models (check our latest paper in the field). Also, in this specific case, MobileODT developed the EVA Portal, where medical practitioners can exchange opinions with their peers, ask questions, etc.
     • Treatment can be done at the point of care, materials are low cost and do not require major technology/infrastructure.
9. What are you doing to increase the interpretability of the classification models that you are using?
   • We have published a couple of papers on this domain to make CNN give explanations with FEUP and INESC TEC. You can see some of them here:
     • https://www.sciencedirect.com/science/article/abs/pii/S0167865520301240
     • https://ieeexplore.ieee.org/abstract/document/8852409
     • https://link.springer.com/chapter/10.1007/978-3-030-31332-6_1
     • https://link.springer.com/chapter/10.1007/978-3-030-02628-8_15
10. What are the major trends in AI consulting right now?
    • The major trend in AI consulting is off-the-shelf models or as I call it, dumbing down the business to the technology. Our focus is right the opposite, elevating the technology to adapt to the business. You can see some examples of how we have done this on one of our previous blog posts.

Projects of interest

If you’re interested in this field and what to know what’s being done in Portugal, I suggest you follow two research projects in the field:

• Project CLARE.
• Project CADPATH.