English Chinese
DHC Research Institute
Artificial Intelligence Image Recognition of Frozen Pathology
Medical Imaging AI Technology

The Medical Imaging AI Technology Department of DHC Research Institute has undertaken a number of scientific research cooperation projects of National Cancer Center, involving the intelligent diagnosis of breast cancer frozen pathology, intelligent delineation of radiotherapy target area and organs at risk, identification and screening of pulmonary nodules by chest CT and other directions, with rich work experience and technical accumulation in image processing, machine learning, neural networks and other aspects.

Technical framework



Competition Data Results

On the Camelyon 2016 competition dataset, the model AUC reached 0.9742.

Clinical Data Results


Screen the images with wrong recognition or poor recognition effect of artificial intelligence

In order to meet the needs of clinical application, it is explored to screen the sections that are easy to identify errors with specific indicators for manual review. The number of sections to be manually reviewed accounted for about 25% of the total, and the correct rate of tumor identification of the remaining sections could reach 100%.


Technology Upgrade - Modeling simulated pathologist films

Referring to the reading process of pathologists, the model innovation is carried out to solve the problems of noise in the original technical framework and poor recognition results for lesions with smaller diameters.


Application Example - Intelligent Identification of Frozen Pathology of Sentinel Lymph Node in Breast Cancer

The intelligent recognition system of frozen pathological images based on the above algorithm has been developed and deployed and launched in Cancer Hospital, Chinese Academy of Medical Sciences, realizing the functions of digital pathological section transmission, intelligent identification of tumor metastases, manual reading of prompts and submission and review.


More application scenarios - multi-example learning without marking the tumor area

Based on tens of thousands of digital sections that only give the final diagnostic result, without the need for doctors to label the tumor area, the trained AI model will be able to automatically identify the tumor area in the section.


Clinical-grade computational pathology using weakly supervised deep learning on whole slide images., Nat. Med., 25, 1301-1309.

More application scenarios - predicting gene mutations from pathological images

The resulting convolutional neural network model, trained using digitized sections of biopsy tissue as raw input data, enables the prediction of many different genetic mutations, and genotype information has been demonstrated to be extractable from histopathological structural patterns.


Classification and mutation prediction from non–small cell lung cancer histopathology images using deep learning. Nat Med. 2018;24:1559-67.

International Cooperation

Our company has a deep collaboration with Philips to meet the current needs of histopathology with the Philips Ultra Fast Scanner high-throughput bright field scanner product. In the field of pathology digitization, the company builds efficient, safe and reliable digital pathology total solutions for users.

Three stages in the development of clinical digital pathology


Pathology and business collaboration


Three elements of pathology AI product development selection


Big data platform concept for pathology images