Intelligent analysis of medical images refers to the use of artificial intelligence technology to identify and analyze medical images, help doctors locate the disease, analyze the condition, and assist in making diagnosis. At present, more than 90% of medical data come from medical images, and most of these data need to be analyzed manually.

The disadvantages of manual analysis are obvious. The first is that it is inaccurate and can only be judged by experience. It is easy to misjudge. According to a misdiagnosis data from the Chinese Medical Association, the number of misdiagnosis in clinical medicine in China is about 57 million every year, with a total misdiagnosis rate of 27.8%, an organ heterotopic misdiagnosis rate of 60%, an average misdiagnosis rate of 40% for malignant tumors, such as nasopharyngeal cancer, leukemia, pancreatic cancer, etc., and an average misdiagnosis rate of more than 40% for extrapulmonary tuberculosis, such as liver tuberculosis, stomach tuberculosis, etc. The second is the big gap. According to the data of the Arterial Network Eggshell Research Institute, the annual growth rate of medical imaging data in China is about 30%, while the annual growth rate of the number of radiologists is about 4.1%, with a gap of 23.9%. The growth of the number of radiologists is far less than the growth of imaging data. This means that the pressure of radiologists to process image data in the future will be increasing, even far beyond the load. If we can use algorithms to automatically analyze images, and then compare images with other case records, we can greatly reduce medical misdiagnosis and help make accurate diagnosis.

The implementation process of AI in the medical imaging industry is roughly as follows: image data preprocessing ->sample cleaning, labeling  model building and training debugging  large-scale data training and verification to obtain a deep learning network model. The above process is the offline training process of AI, and the final output is the deep learning model. Then the generated model can be used for online prediction or auxiliary judgment.

Solution introduction

Provide end-to-end AI solutions for medical imaging, as shown in the figure below, and realize the following three functions.

(1) Sample data preprocessing. Each laboratory of the hospital, such as CT, BT, CR, etc., transfers medical image data from the terminal equipment to the parallel storage through the 10-gigabit/IB network. The data preprocessing CPU platform (a cluster of multiple dual-channel CPU servers NF5280M5) reads the data from the storage, runs the edge detection segmentation, region growth segmentation, seed algorithm and other programs, obtains the target data, and then labels it to form a training sample library, which is stored in the parallel storage. The management, scheduling and monitoring of CPU programs will be completed by the unified management platform AIStation.

(2) Model training. The model training GPU cluster (configured with a stand-alone 8-card GPU server, such as NF5288M5) will read the training sample library data from the parallel storage, load the CNN model, run the in-depth learning framework, such as TensorFlow, Cafe, Mxnet, etc., to train the initial model, and generate the final model after learning and training a large number of data samples. The training involves the submission of multiple training tasks, and its resource management, scheduling and monitoring will be completed by the unified management platform AIStation.