In-Depth Imaging: A Pathology Diagnosis System Based on TensorFlow

By / Wang Shuhao

According to the World Health Organization (WHO), malignant tumors are the second leading cause of death globally, causing nearly ten million deaths each year. The diagnosis of malignant tumors requires sufficient evidence, with histopathological diagnosis being the most reliable method for tumor diagnosis, serving as the fundamental basis for cancer confirmation and treatment. The diagnosis made by pathologists is regarded as the ultimate judgment of the condition.

Pathology is the gold standard of medical diagnosis, and pathology reports play a crucial role for clinicians in further diagnosing and treating diseases. Currently, there are only about 10,000 professional pathologists in China. According to the former Ministry of Health’s document No. 31, there should be at least one to two pathologists for every hundred beds. Based on this standard, there is a shortage of at least 50,000 to 60,000 pathologists in China. The long training period for pathologists and the talent gap have led to overwork.

Comparison of Digital Pathology Slides and Radiographic Images

Traditional pathology departments use microscopes to observe physical slides and obtain diagnostic reports. With the development of digital pathology, an increasing number of slides have been digitized, making computer-aided diagnosis possible. Unlike CT, X-rays, and other radiographic images, the volume of pathological images usually ranges from 500 MB to 1 GB, and they contain a large amount of information, posing significant challenges for the construction of deep learning models and the architecture of distributed computing systems. The AI pathology assistance diagnosis system, ThoroughInsights, based on the TensorFlow ecosystem, brings groundbreaking innovations to the pathology industry.

Data, algorithms, and systems constitute the three pillars of AI product development.

Digital Pathology Slide Annotation Tool ThoroughWisdom

Data is an important entry point for intelligence. We collaborated with over 40 top-tier hospital pathologists with more than 10 years of experience to collect thousands of high-precision annotated digital pathology slides using our self-developed annotation tool, ThoroughWisdom, based on iPad and Apple Pencil, covering multiple organs of the digestive and respiratory systems. For example, in the gastric model, after segmentation, the number of training images exceeds 50 million, making it the largest high-quality pixel-level annotated dataset in the industry.

Deep Learning Model Training Pipeline

With training data in hand, the next step is to establish the deep learning model. During the product development process, we went through three stages: customization based on existing models, self-developed models, and automated machine learning (AutoML), gradually improving the model’s accuracy and generalization ability. Based on TensorFlow, we established a complete deep learning pipeline. On a deep learning cluster with 10 GPU cards, when the training data reaches a scale of 50 million, a model iteration can be completed within half a week. According to the test results of nearly 10,000 gastric digital pathology slides, the gastric cancer recognition model in ThoroughInsights can achieve 100% slice-level sensitivity and over 85% specificity.

Distributed Microservice Inference Architecture

After model training is complete, we built a distributed microservice inference architecture based on TensorFlow Serving. The system has high availability and scalability, allowing prediction tasks to be distributed across all GPUs in the cluster for parallel computing, significantly increasing the speed of digital pathology slide predictions.

Pathology Assistance Diagnosis System ThoroughInsights

Currently, the gastrointestinal artificial intelligence pathology assistance diagnosis system has been implemented in five large tertiary hospitals, making it the world’s first large-scale deployed histopathological assistance diagnosis system.

During the diagnostic process, pathologists can not only diagnose diseases through monitors but also refer to the computer’s auxiliary analysis results simultaneously. The AI-assisted diagnosis system has advantages such as tirelessness, objective accuracy, and efficient computation, which can help doctors filter out a large number of negative cases, significantly reducing the workload of pathologists.

Pathologists can delegate some repetitive and cumbersome pathological diagnosis tasks to the assistance diagnosis system, such as gastrointestinal pathology. This allows pathologists to have more time for the diagnosis and treatment of complex diseases and research in cutting-edge fields, promoting the advancement of medical technology.

More Cases:

• TensorFlow Lite + AoE: A Deep Learning-Based Road to Driving Safety

• TensorFlow Helps “Safe Good Translation” Win the Championship in WMT 2019 “English to Chinese” Track

• Human Action Recognition in the Smart Eight-Section App

• Providing Preset Suggestions for Images: Building the “Photo Dedicated” Feature in VSCO

• Kingsoft WPS: Document Correction Function Based on Local TensorFlow Lite and AI Cloud Services

• Ecovacs Robotics: AI Vacuum Cleaner Collaborates with TensorFlow to Explore Indoor Scenes

• Case Sharing | iQIYI’s Mobile AR Solution SmileAR Based on TensorFlow Lite

• China Telecom Uses TensorFlow to Achieve Artificial Intelligence

• Ask Me Out: Using TensorFlow Lite to Deploy Hotword Detection Models on Embedded Systems

• Homework Practice with TensorFlow Lite

• When TensorFlow Meets Youdao Translation King 2.0 Pro

• Zhejiang Mobile: Using TensorFlow to Achieve Intelligent Communication Operations

• Shanbay: Applying TensorFlow for Deep Knowledge Tracking

• Writing Assistant: High School Students Use TensorFlow to Promote Chinese Character Dissemination

• Kingsoft WPS: AI Mobile Office Based on TensorFlow

(Click “Read Original” to submit a case, and we will contact you as soon as possible)