Winning Health and Intel collaborated to accelerate healthcare solution and assist in enhancing patient care.
Overview
Winning Health is an AI healthcare technology company that helps hospitals get the benefit of deep learning solutions. This project uses the 4th generation Intel® Xeon® Scalable processor with Intel® Extension for PyTorch* to optimize an AI 3D medical visualization solution based on segmentation.
In recent years, with the continuous advancement of AI healthcare technology, 3D segmentation techniques have been applied in the field of oral medicine. This not only provides patients with an intuitive and visual representation of their oral cavity, allowing them to gain a comprehensive understanding of the details of their individual oral structures, but also enables doctors to visually observe oral structures during procedures. This aids in the formulation of more precise treatment plans.
Winning Health has achieved oral 3D visualization based on Cone Beam Computed Tomography (CBCT), accurately segmenting individual teeth along with root information. This plays a significant role in assisting doctors, particularly in procedures such as dental implant treatments.
Algorithms
With the support of deep learning technology, in contrast to traditional manual segmentation, a neural network trained on real data can automatically identify and segment key areas in Cone Beam Computed Tomography (CBCT) without the need for human intervention. For doctors, tasks that are challenging or time-consuming in segmentation can be rapidly and accurately accomplished by AI models. This is a great assistance to the doctors. Currently, research in this field is increasing rapidly, and Winning Health engineers have addressed the aforementioned challenges using nnU-Net, a mature, open source 3D medical image segmentation framework.
The specific algorithmic process is shown in Figure 1. The CBCT of the teeth undergoes a series of preprocessing steps and is fed into a deep neural network model to obtain masks for the dental arch and dental nerve canal. After post-processing these two results, they are combined to generate the final 3D segmentation result.
Figure 1. Pipeline of 3D tooth segmentation algorithms
Challenge
When implementing and deploying algorithms in medical institutions, Winning Health encountered several challenges, including:
- The complexity of hospital business scenarios with high requirements for data confidentiality, security, and real-time processing. The application cannot be connected to the internet and needs to be deployed within the internal network. Additionally, it must comply with the institution's data protection policies and adhere to relevant privacy regulations.
- The segmentation algorithm places specific demands on the computing and memory resources of medical institution equipment. Beyond necessitating substantial computing power for inference and result generation, the equipment must possess ample memory to accommodate both the storage of models and large CBCT files.
- The CBCT file size is considerable, necessitating a patch-based iterative inference process followed by result integration. Consequently, this results in extended forward times.
To address these challenges and enhance the inference performance and efficiency of the 3D segmentation algorithm on CPU-based medical institution devices, Winning Health collaborated with Intel in a joint research effort. Together, they conducted a series of performance optimizations for the algorithm.
Solutions and Technologies
The joint optimization mainly involves the following aspects:
- Using Intel® Advanced Matrix Extensions (Intel® AMX) to enhance the performance of tooth 3D segmentation on the CPU to optimize the entire pipeline from data input and deep learning to the inference process.
- Using graph optimization to reduce the overhead of operator and kernel calls. Different computation methods may exhibit significant performance differences at the inference engine and hardware level. Graph optimization is an optimization technique for deep learning models, reducing the overhead of operator and kernel calls through operator fusion, algorithm adjustments, and other methods to improve model runtime efficiency.
- Employing multi-instances to fully use the performance of CPU multicores. Intel Xeon Scalable processors, known for their high computing power and numerous cores, benefit from multi-instance parallel execution to effectively enhance CPU use and achieve better throughput.
Performance Results and Benefits
Through the collaborative optimization efforts of Intel and Winning Health, a cost-effective delivery model for the 3D tooth segmentation model was established. By refining algorithms for key operators in PyTorch on the CPU platform, the deep learning framework's inference speed was further accelerated.
In the experiments, the joint optimization team used Intel Extension for PyTorch and conducted BF16 optimization performance comparisons for different models of the 4th generation of Intel Xeon Scalable processors. From the results data, we can observe that with the use of Intel optimizations, there has been an improvement in the performance of workloads. Compared to the 3rd and 4th generations of Intel Xeon Scalable processors, we can also conclude that the half-precision optimization of the 4th generation of Intel Xeon Scalable processors significantly enhances performance.
The main optimization behind the 4th generation of Intel Xeon Scalable processors is Intel AMX. It is a dedicated instruction set and matrix multiplication accelerator introduced by Intel for the 4th generation of Intel Xeon processors. It consists of two components the TILE and the Tile Matrix Multiply (TMUL), an acceleration engine connected to TILE that performs matrix multiplication calculations for AI. With this architecture, Intel AMX enables faster processing of matrix multiply-add operations for BF16 or int8 data types, significantly improving AI applications.
Figure 2. Relative speedup for 3D tooth segmentation workloads on Intel Xeon processor 8358 versus Intel Xeon processor 6438y with different optimizations
Summary and Conclusion
Winning Health has developed a 3D Tooth Segmentation solution. Using Intel Extension for PyTorch has facilitated the deployment of this solution on Intel Xeon Scalable processors for efficient inference. As a result, a high-resolution CBCT can be converted into a 3D model in just two minutes, rivaling alternative GPUs available in the market. This level of performance ensures that our solution not only meets but exceeds our high standards for efficiency and effectiveness, delivering a superior end-user experience.
About Winning Health
Founded in 1994, Winning Health Technology Group Co., Ltd. is on a mission to "empower with technology, elevate people's health." The company's business covers smart hospitals, intelligent regional health, internet plus medical health, and more, aiming to become a trusted service provider in the digital health field.
The company's headquarters is located in Shanghai, with 10 research and development bases and 20 branch offices nationwide. It serves over 6,000 medical and health institutions, including more than 400 tertiary hospitals.
Winning Health AI Lab
Established in 2016, the lab is composed of members from renowned domestic and international universities, with 90% holding master's degrees or above. It brings together professionals from interdisciplinary fields such as computer science, medicine, and statistics.
Research Focus
The lab focuses on exploring AI in the field of medical health. It actively collaborates with domestic and international medical institutions, universities, and research institutions to explore innovative technological research systems. Research directions include medical big data analysis, natural language processing, computer vision processing, and more.
Honors and Achievements
- Eight patents granted
- First prize in the Shanghai Science and Technology Award for Technical Invention
- Third prize in the Pudong New Area Science and Technology Progress Award
- In 2022, CHIP2022: First place in "Gene-Disease" Association Semantic Mining; First place in Extracting Diagnosis and Treatment Decision Trees from Medical Text
- In 2021, CHIP2021: Second place in the Medical Dialogue Clinical Discovery Positive and Negative Discrimination Task
- In September 2021, the 20th Chinese Computational Linguistics Conference: First prize in the Technical Evaluation Task - Intelligent Medical Dialogue Diagnosis and Treatment Evaluation: Medical-Patient Dialogue Understanding.