Blood tests tell doctors about the function of key organs, and can reveal countless medical conditions, including heart disease, anemia and cancer. At major hospitals, the number of blood cell images awaiting analysis can be overwhelming.
With over 10,000 beds and more than 8 million outpatient visits annually, Taiwan’s Chang Gung Memorial Hospital collects at least a million blood cell images each year. Its clinicians must be on hand 24/7, since blood analysis is key in the emergency department. To improve its efficiency and accuracy, the health care network — with seven hospitals across the island — is adopting deep learning tools developed on AI-Ready Infrastructure, or AIRI.
An integrated architecture from Pure Storage and NVIDIA, AIRI is based on the NVIDIA DGX POD reference design and powered by NVIDIA DGX-1 in combination with Pure Storage FlashBlade. The hospital’s AIRI solution is equipped with four NVIDIA DGX-1 systems, delivering over one petaflop of AI compute performance per system. Each DGX-1 integrates eight of the world’s fastest data center accelerators: the NVIDIA V100 Tensor Core GPU.
Chang Gung Memorial’s current blood cell analysis tools are capable of automatically identifying five main types of white blood cells, but still require doctors to manually identify other cell types, a time-consuming and expensive process.
Its deep learning model provides a more thorough analysis, classifying 18 types of blood cells from microscopy images with 99 percent accuracy. Having an AI tool that identifies a wide variety of blood cells also boosts doctors’ abilities to classify rare cell types, improving disease diagnosis. Using AI can help reduce clinician workloads without compromising on test quality.
To accelerate the training and inference of its deep learning models, the hospital relies on the integrated infrastructure design of AIRI, which incorporates best practices for compute, networking, storage, power and cooling.
AI Runs in This Hospital’s Blood
After a patient has blood drawn, Chang Gung Memorial uses automated tools to sample the blood, smear it on a glass microscope slide and stain it, so that red blood cells, white blood cells and platelets can be examined. The machine then captures an image of the slide, known as a blood film, so it can be analyzed by algorithms.
Using transfer learning, the hospital trained its convolutional neural networks on a dataset of more than 60,000 blood cell images on AIRI.
The AI takes just two seconds to interpret a set of 25 images using a server of NVIDIA T4 GPUs for inference — a task that’s more than a hundred times faster than the usual procedure involving a team of three medical experts spending up to five minutes.
In addition to providing faster blood test results, deep learning can reduce physician fatigue and enhance the quality of blood cell analysis.
“AI will improve the whole medical diagnosis process, especially the doctor-patient relationship, by solving two key problems: time constraints and human resource costs,” said Chang-Fu Kuo, director of the hospital’s Center for Artificial Intelligence in Medicine.
Some blood cell types are very rare, leading to an imbalance in the training dataset. To augment the number of example images for rare cell types and to improve the model’s performance, the researchers are experimenting with generative adversarial networks, or GANs.
The hospital is also using AIRI for fracture image identification, genomics and immunofluorescence projects. While the current AI tools focus on identifying medical conditions, future applications could be used for disease prediction.