第14回 Spark Labsレポート

Upon arrival at the San Jose McEnery Convention Center, there was a big sign that had the words: ‘GPU Technology Conference’ written. As soon as you entered the building, you were welcomed by many people wearing NVIDIA shirts or green striped neck straps to indicate that they were staff members. Surprisingly there were no lines/queues at the reception so getting in was no problem.

Later in the afternoon, at the gear store, apparels and many other items were being sold. However what interested people most were the GPUs, systems, devices, chips and other tech accessories. Especially the NVIDIA Jetson TX2 Developer Kit. This small module has the capability to be used for AI and includes AI algorithms/libraries thanks to the NVIDIA Pascal architecture. The size allows it to be used in edge devices such as robots, drones, smart cameras, medical devices etc. We are now using these for research purposes at Azest.

The roof at GTC was not very high but the hall was very long and was filled with posters highlighting the work of companies and universities. These posters resembled scientific papers but was more visually appealing because of the different background colors and pictures (A company even set up a beer and poster event where you could get a free beer if you read their poster!). Some notable works were by NVIDIA and John Hopkins University on Convolutional Neural Network auto-encoding and machine learning medical imaging focused on endoscopy data.

The keynote event started in the morning of the 2nd day at 09:00. Upon arrival, the queue was very long and Katori san had to wait 1 hour and 30 minutes to get in. He managed to get in by 08:30 however others who arrived at 08:30 or later had to stand or could not even get in because of the lack of space.

NVIDIA welcomed everybody and started their presentation about the GPU computing evolution/history of GPU and how it had evolved since then. Then continued to talk about industry adoption - i.e. what industries have been using NVIDIA to improve their work such as manufacturing (e.g. FANUC) or social (e.g. Jibo - Robot).

In order to highlight NVIDIA’s computing power and their case for supercharged computers, a presentation about CLARA, a medical imaging supercomputer was shown. This research was conducted by Harvard Medical School, GE & many other related groups. They are trying to reduce heart disease (Shows a video of real time Cardiac Echo/Echocardiography/心エコー). This supercomputer can create automated annotations around the organs and automatically detect whether the doppler ultrasound taken is inadequate. They have started collecting data and are planning to release to different regions in about a year’s time. When the supercomputer runs the algorithms, it makes a lot of data because it creates many multimodal data.

Dr. Rubin’s presentation was about deep learning with brain surgery, lung cancer and many others. Next, a slide on image segmentation with lung cancer was shown: The images are taken using CT, MRI, Ultrasound and then with deep learning, are segmented: it divides the areas of the lung into non-overlapping homogenous regions by coloring each area of the lung and highlight the irregularity (e.g. feature extraction, image classification).

The next slide was about image segmentation in pathology. AI will detect E.g. Nodules/areas of concern in cell/nuclear regions or tissue regions and segment with different image sizes/shapes. When looking at the shapes of these areas, AI outputs the mathematical functions and compares/matches the segmented images and pathology images (used as ground truth). Doing this allows for analysis. If hospitals and universities in Japan don’t do this, they will not be able to do any meaningful analysis.

Katori san noticed that during the whole conference, no one mentioned about how AI can do diagnosis. There are many instances where AI can support diagnosis but in reality, AI has a hard time to diagnose. We are still in early stages but are trying to prepare for the next stage.

Dr. Kumar’s main focus was on breast cancer but now he started to focus more on Thyroid. His presentation during GTC was about breast cancer and the ultrasound imaging process on how to get to the end image (Transducer -> Raw Channel/RF data -> Beam formed RF data -> In-phase quadrature data -> B-mode image -> post processed B-mode image -> end).

The difference between the expert’s (clinician) segmented image and the AI predicted segmentation image was 0.94. Overall it was 0.82. Overall, segmentation and classification can help to reduce localization time, help sonographers in categorizing suspicious nodules and be used in real-time.

In conclusion, GTC was an exhilarating and valuable experience. Especially the opportunity to see how NVIDIA has been a considerable factor in many diversified environments as well as being able to see the innovative progress in the medical field.

By : Mac

SparkLabs