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.

In recent years, with the continuous development of AI technology, AI systems based on deep learning algorithms mainly focus on the analysis and research of mammographic X-ray images, and there are relatively few studies on MRI, ultrasound and digital breast tomography. Some domestic medical companies have also introduced a new generation of molybdenum target AI system based on deep learning algorithm, which can achieve more than 90% accuracy in breast mass detection and calcification detection, almost equivalent to the level of medical imaging experts. At the same time, in the discrimination of benign and malignant lesions, the new generation of AI model can achieve 87% sensitivity and more than 90% specificity, even beyond the level of medical imaging experts.

Although the AI system for mammography has been tested to have good diagnostic performance, it lacks in-depth research on breast imaging application scenarios. Algorithm engineers fail to effectively excavate the pain points and difficulties in clinical diagnosis, making the breast molybdenum target imaging AI system unable to accurately reflect the real needs of imaging physicians.

In recent years, the incidence and mortality rate of lung cancer has ranked first among the common malignant tumors in China. Early-stage lung cancer is mostly manifested as lung nodules, which are small in size, low in contrast and highly heterogeneous in shape. Therefore, early detection, early prevention, early diagnosis and early treatment can reduce the incidence of lung cancer to a great extent. An important tool in lung cancer prevention and treatment is early screening, of which low-dose CT of the chest is an internationally recognized and effective tool. However, with the increasing number of people being screened by CT chest, the workload of imaging physicians is increasing. The heavy, tedious work of reviewing films increases the fatigue of imaging physicians, as well as the risk of missed and misdiagnosis.

In China, the medical field is one of the relatively booming areas for AI development. AI applications based on deep learning now cover all clinical stages, including lesion detection, pathological diagnosis, radiotherapy planning and post-operative prediction, etc. The application of AI algorithmic models has greatly reduced the workload of imaging physicians. At present, several tertiary hospitals have collaborated to develop various AI models for lung nodules and have applied them in their clinical work, all with good results.

Although deep neural networks have been initially validated in terms of their effectiveness in diagnostic problems, the clinical work in medical imaging is often intertwined with a variety of different tasks. It is easy to see from the screening and diagnostic aspects of lung nodules that AI needs to play a reliable role in the detection of abnormalities, quantitative measurement,  follow-up and differential diagnosis before it can ultimately benefit the clinical application. In addition to applying deep neural networks to the classification problem (diagnosis) of medical images, researchers need to continue to explore the application of AI techniques to the medical image detection problem (abnormality detection), segmentation problem (quantitative measurement) and alignment problem (follow-up tracking).

The auxiliary diagnostic system of mammography is based on mammography, which realizes the diagnosis and treatment of breast diseases through the close cooperation of "AI + medical treatment". This protocol is completely in accordance with the daily working mode of imaging physicians. Medical imaging experts accurately identify and label all breast masses, calcifications, structural distortions and other characteristics on the images according to the experience accumulation and work summary in the hospital for many years. The program makes a complete description of the shape, size, density, and nature of breast hyperplasia, lesions, lesions, and calcifications, and makes important decision preparations for clinical intervention, which greatly improves the work efficiency of imaging physicians and reduces the incidence of missed diagnosis and misdiagnosis.

The Lung Nodule Image Assisted Diagnosis System is based on CT imaging of the chest, and uses the close collaboration of "AI+Medicine" to diagnose and treat lung diseases. The program uses a deep learning neural network to segment the annotated lung nodules in accordance with the daily work pattern of imaging physicians, and medical imaging experts identify all the lung nodules on the image based on their years of experience and work in hospitals. The solution also provides a complete description of the size, density, voxel and morphology of the nodules, determines the growth pattern and nodule type, and assist physicians to make important decisions for clinical intervention, greatly improving the efficiency of the imaging physicians and reducing the incidence of missed and misdiagnosis.

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

1.  Automatically parses patient and image details based on uploaded image resources.

2.  Provide query function based on fields such as device type, examination site, impact number, examination time period and upload time period of the image.

3.  Adjustments can be made to the film's greyscale, window width window position, etc. as required

4.  Gradual interpretation, with the possibility of marking the position of nodules and remarks information using different labelling tools.

5. The size, density, voxel and other important information of each lesion can be calculated automatically, and the physician can make an accurate judgement on the morphology, nature and location of each lesion.

6. Statistics on suspected lesions.

7. Record the uploaded images and their detailed information and lesion information into the library to provide support for later lesion follow up and clinical diagnosis.

1.  Automatically parses patient and image details based on uploaded image resources.

2.  Provide query function based on fields such as device type, examination site, impact number, examination time period and upload time period of the image.

3.  Adjustments can be made to the film's greyscale, window width window position, etc. as required

4. Gradual interpretation, the location of the nodules and the remark information can be marked using different marking tools.

5.  Automatic calculation of the size, density, voxel and other important information of each nodule, the physician can make an accurate judgment on the morphology, nature and location of each nodule.

6.  Statistics on suspected lesions.

7.  Record the uploaded images and details as well as the lesion information into the library to provide support for later lesion follow up and clinical diagnosis.