Taking deep learning to the next level
Notes from Aiforia at the Digital Pathology and AI Congress 2019
We had high expectations for an engaging and stimulating few days at the ‘5th Annual Digital Pathology and AI congress’ in New York last week. The experience was almost overwhelming, we can’t recall a meeting with more lively discussions. We found ourselves spending less time pitching the possibilities of Deep Learning for digital pathology, and more time proposing strategies for solving specific end user Image Analysis (IA) problems, ranging from context dependent quantitation of IHC markers to enable separation of normal and pathological structures, to complex spatial relationship analysis in high multiplex immunofluorescence stacks.
We are grateful to our two invited speakers, Dr. Jean-Martin Lapointe from AstraZeneca and Dr. Gillian Beamer from Tufts University, for taking the time to join us in New York and present their work. The discussion during the fully-packed workshop was riveting. There was a noticeable shift from an AI curious crowd to a knowledgeable one, focusing on intended use, regulatory aspects, and validation strategies. A big thanks as well, to all attendees for making it a success.
During our hands-on segment of the workshop we saw more than 30 users with their own laptops, some even on tablets, log in to our cloud server and simultaneously annotate and train their own breast cancer AI models. The ability to do this at a practical level is a testimony to the robustness of the Aiforia platform. We are deployed as a browser GUI, scalable cloud solution (Azure) using state of the art GPUs (NVIDIA), and can support an unlimited number of users, at anytime, anywhere in the world.
All computation is performed in the cloud and requires no IT investments on the part of our users. They were connected over two standard wireless hotspot devices, and for the 1.5 hours of the hands-on workshop we generated less than 3GB of network traffic combined. This means if you have a standard mobile phone data connection you will have a world class data center, computational resources, and AI infrastructure at your fingertips.
The task was to generate a multilayered, multiclass AI model for i) quality filter for in focus tissue ii) epithelia vs stroma segmentation and iii) Ki67 IHC positive vs negative counts for specific compartments. We believe that everyone walked away, as what we at Aiforia like to call, #AiReady.
Three take home messages from the conference:
- If you need to insert a computational scientist, data scientist or IA software expert between the problem and the pathologist you have already failed.
- Unless you can turn on any computer you already have at hand and readily access the needed computational resources, without need for IT stakeholders, command line or other advance data interfaces, you have already failed.
- If you are unable to adjust your AI model to a growing and scalable ground truth you have already failed.
Our solution aims to deliver on all of these pain points and an end-to-end solution for the user. With this in hand the pathologist, clinician, or researcher can apply their data and robust domain expertise to supervise the AI training process and take a leap forward in how they approach questions previously untenable. Most importantly, the decision-making process will now rest in their hands. Our mission is to democratize AI in digital pathology, to give the decision-making power to the person that is the expert.
If you are ready to take the next steps and see the platform in action with your data, We can take up to 5 images and mock up an AI model for you and show you your own neural network.