Breaking the international record of AI-Assisted Thoracic Multi-Organ Segmentation, Tencent Youtu won three championships at SegTHOR.


New Breakthrough in Global AI-Assisted Thoracic Multi-Organ Segmentation Technology. Recently, a world record in 2019 Segmentation of Thoracic Organs at Risk in CT Images was once again refreshed. TencentX team, which is composed of Tencent's top AI laboratory-Tencent youtu laboratory and Wang liansheng laboratory of Xiamen university, outperformed over 600 scientific research teams, achieving three first and one second prizes in the most important Dice index (a similarity coefficient, which is mainly used to indicate the similarity between model output and real distribution), with the accuracy reaching the highest level in the world.


Tencent Youtu TencentX team won three championships at SegTHOR

SegTHOR Challenge 2019, co-sponsored by CodaLab and ISBI, is dedicated to solving the problem of organ risk segmentation in computed tomography images, helping doctors to improve manual rendering efficiency and reduce anatomical errors in clinical treatment. The competition attracted 638 teams from around the world to register for participation.

According to the latest data released by the World Health Organization, the number of cancer patients worldwide increased by 18.1 million in 2018, with 9.6 million people dying of cancer. Worldwide, one in 8 to 11 people die of cancer. Radiotherapy is not uncommon in the treatment of cancer. However, radiotherapy may damage healthy cells while killing cancer cells. Therefore, the accurate separation of healthy organs near the tumor of patients will directly affect the rehabilitation of patients.

At present, organ delineation based on three-dimensional CT images commonly used in clinic relies heavily on manual labeling by experts. Doctors need to label multiple organs continuously, intensively, and simultaneously, which is not only heavy in workload and low in efficiency, but also takes up to 96 minutes to delineate tumor regions of a single patient according to different diseased positions. In addition, the target organs of different patients usually differ greatly in shape and position, and the contrast of contours in CT images is relatively low, which also brings difficulties to the segmentation and labeling of organs.


Chest CT slice, left is CT image, right is the result of segmentation. The green area corresponds to the heart, the yellow area to the aorta, the blue area to the trachea, and the red area to the esophagus. 

The thoracic multi-organ segmentation system invented by Tencent Youtu TencentX team provides a fully automated healthy organ segmentation method based on convolution neural network structure, which can accurately and quickly segment healthy organs located near target tumors in thoracic cavity. 40 groups of CT images are trained, which is realized in two stages: fast positioning module and fine segmentation network. Firstly, fast positioning is realized by reducing the spatial distance sampling of 3D CT images and matching with 3D convolution neural network. Then, the 2.5D convolution neural network is used to predict the results layer by layer along the sagittal plane, coronal plane and horizontal plane. At the same time, the results are combined with the results directly predicted by the 3D convolution neural network under the complete VOI to generate more accurate and comprehensive multi-organ segmentation results for the purpose of fine segmentation. This first-positioning-then-segmentation approach not only eliminates a lot of interference for the final segmentation, but also saves computing resources.


Logic Scheme of Tencent Youtu TencentX Team's Chest Multi-organ Segmentation System

The thoracic multi-organ segmentation system can not only help doctors to control the dose more accurately in radiotherapy and liberate them from labor-intensive organ labeling, but also improve the segmentation accuracy to guide doctors to give patients more accurate and quick treatment. It effectively reduces the treatment risk, and also helps medical researchers to better understand and study thoracic organs.

Tencent Youtu has all along been committed to medical AI. Not long ago, it just set two records in the global medical imaging competition LiTS, wining two world's first prizes in liver segmentation and liver tumor segmentation. By deepening its technological R & D, Tencent Youtu has launched its first medical imaging product "Tencent Minying". Currently, this system has supported cancer screening for cervical cancer, lung cancer, ophthalmic diseases and other cancers, landing in more than 100 top 3A hospitals in China. It not only reduces the workload of doctors, but also plays an important role in improving diagnostic accuracy and efficiency.