AI essential medical floor! Tencent excellent view of the industry's first open source 3D medical image data of a large pre-training model

640?wx_fmt=png

Recently, Tencent excellent view of the first pre-medical AI depth learning training model MedicalNet officially open. This is the first global project to provide medical imaging dedicated pre-training model of a variety of 3D, AI will provide the basis for global health development.

Many studies have shown that the development of deep learning is very dependent on the amount of data. Natural image area, there are many large data sets, such as ImageNet, MSCOCO. Pre-training model based on these data sets produced to promote the classification, testing, progress segmentation applications. Most of them are different from the 3D structure and morphology of natural images, medical imaging, and at the same time, because the data acquisition and labeling difficult, sparse amount of data, large data sets and the corresponding pre-training model not yet exist.

MedicalNet (https://github.com/Tencent/MedicalNet) Tencent excellent view of the first dedicated 3D medical imaging in depth learning application developed by a series of pre-training model for any 3D medical imaging applications of AI to play "hit the role of the foundation "to speed up the convergence of the model, the model to reduce dependence on the amount of data, MedicalNet have the following characteristics:

1. MedicalNet pretraining network can migrate to provide any AI 3D medical imaging applications, including but not limited to, segmentation, detecting, classifying and other tasks;

2. especially for small medical imaging data AI scene, can speed up the convergence of the network, improve network performance;

3. Configure a small value through a simple interface parameters, to fine-tune the training;

4. The project provides training and testing multi-card evaluate code, interface, rich, and strong expansion;

5. Provide different depths 3D ResNet pre-training model, the order for the application to use different data.

In order to produce a pre-training model of 3D medical images, MedicalNet gather more semantic 3D from different medical fields divided small data sets, and a multi-domain multi-branch joint training model based decoder to solve the problem of lack of tagging data set. Our pre-training model can migrate to deep learning model any 3D medical imaging applications. Workflow of the whole system as shown below:

640?wx_fmt=png

We have not talked to migrate MedicalNet model to the pre-training Visceral and LIDC dataset, complete the new segmentation of the lungs and pulmonary nodules classification tasks, and with the commonly used from zero training (train from scratch) and 3D video pre-Kinetics training model in performance and convergence speed compared. Segmentation applications in the lungs compared to the Train from Scratch, MedicalNet has been raised 16% to 33% of the amplitude in the Dice, Kinetics has been raised as compared to 4% to 7% of the amplitude. On the application of classification of benign and malignant pulmonary nodules, as compared to the Train from Scratch, MedicalNet 6% to 23% of the amplitude of the prediction accuracy (Acc) increase, as compared to 7% Kinetics has been raised to 20% of the amplitude.

On the convergence rate, experiments show, either in the lung or pulmonary nodule segmentation task classification tasks, MedicalNet can provide a lower value for the model initialization loss, loss rate of decline has accelerated noticeably, MedicalNet below shows the performance of a simple example, showing the whole organ segmentation application, different training methods pre-test results in a certain number of training iterations. As can be seen, based on the results of our pre-training model (MedicalNet) closest to the label (ground truth), and far better than the results from zero training (train from scratch), and for more details please refer to the paper ( Chen, Sihong, Kai mA, and Yefeng Zheng. "Med3D: 3D Medical Image Transfer Learning for the Analysis." arXiv arXiv Preprint: 1904.00625 (2019). ) .

640?wx_fmt=gif

With the artificial intelligence of the fiery, medical imaging, AI has become the current various applications of the most popular sections. Different from other applications of artificial intelligence, artificial intelligence applications in the medical field highest threshold, the biggest reason is the lack of annotation data. Acquiring medical image data usually needs to undergo a series of hurdles, at the same time, due to the specific areas of data usually requires a physician experienced marked, and labeled each 3D data is time-consuming. In the current tight health care resources, access to medical image data front will be very long, greatly hindered the application of the landing process. Moreover, labeled amount of data is quite limited, most of the health sector are scarce and the amount of data that needs to face the gap between deep learning.

The exclusive domain of proprietary models, MedicalNet equivalent prepared a database with clinical common knowledge for each 3D medical imaging applications. Even a small amount of data, the effective features of the database application can also help to achieve better landing performance of medical testing, which greatly reduce the medical imaging applications relying on the AI amount of data, realized the landing demand, speed up the landing speed.

MedicalNet Tencent in the medical field of AI's first open source project, follow-up will continue to offer more types of models based on building a global healthcare AI booster.

As of August 2019, Tencent has been published on Github 81 open source projects including Tencent AI, micro-letters, Tencent cloud, Tencent security and other related areas, and a total of at Github received more than 230,000 the number of Star, both domestically and internationally harvest attention and recognition.