How to use AI to locate pulmonary nodules in Shanghai Jiaotong University Artificial Intelligence Laboratory
Release date: 2017-10-31
Cancer, like the devil in the dark, brings fear and despair. Lung cancer, as a type of cancer with the highest morbidity and mortality in China, has harmed countless families. Nearly 600,000 people die of lung cancer every year in China. However, the mortality rate of cancer is closely related to the time when cancer was first discovered. Early lung nodule screening can save thousands of people from pain and suffering. Professor Xu Wei, Prof. Ni Bingbing, Professor Yang Xiaokang, and Zhu Ximeng from Shanghai Jiaotong University have cooperated with the technology in the point. They use the deep learning to establish a lung nodule automatic positioning screening system, which can effectively detect lung CT. The image contains various nodules such as tiny nodules and ground glass, and reduces the occurrence of false positive misdiagnosis, realizing the desire of "early detection, early diagnosis, early treatment, early recovery". The algorithm won the first place in the Tianchi Big Data Competition. This competition attracted more than 2,000 teams from all over the country. The total prize pool is up to one million, and it is attended by major hospitals, universities, companies and research institutes.
The team used convolutional neural network technology in the field of computer vision to solve lung nodule detection problems and innovate on multiple levels. 1) Combine the object detection and segmentation algorithm to extract candidate nodules, and generate a high recall rate candidate nodule pool. 2) Use a false positive attenuation network and adopt a multi-scale integrated learning network model to improve detection accuracy and attenuate the false positive ratio. 3) In the processing of data, data generation is generated by generating a confrontation network, which improves the effectiveness of training.
Algorithm framework
Data preprocessing
Data diversity is augmented by a geometric transformation such as rotational translation for a limited number of positive samples, similar to doctors analyzing nodule regions through different perspectives and different contexts, and generating a confrontation network (GAN) from random noise. The new nodule is a positive sample, learning to generate new forms of nodule samples, deepening the diversity of data, and improving the generalization ability of the model.
Nodule pretest
Establish a 3D-Unet network structure. The main function of the segmentation network is to extract suspected candidate nodules, maximize sensitivity, and reduce missed detection rate. The network inputs three dimensions of data features, which can be "observed" from multiple Z-axis dimensions. Just as doctors combine multiple planes to observe nodules, they can fully understand the difference between normal and abnormal textures inside the lungs, capturing multiple nodules. Sexual characteristics, such as the density of frosted glass nodules is slightly higher than the surrounding, cloud-like, the solid density of pure solid nodules is higher, similar to the separate egg yolk.
Nodule detection
Nodule precision detection uses three models to predict the probability of candidate nodules separately, and gives the final probability according to the weight ratio between models. The main advantage is that the negative sample experiences an easy to difficult learning process, and the segmentation network and the subsequent false positive attenuation network complement each other. Multi-structure type model Ensemble, and a single network performance is good, similar to the process of multiple doctors reading the film independently, giving a comprehensive reading result.
result
This algorithm won a total victory in the Tianchi Medical AI Competition jointly organized by Alibaba Cloud and Intel. It stood out from the more than 2,000 teams with a score of 0.732 and ranked first in the most important rematch of the competition.
The model trained according to the algorithm can better deal with different forms of nodule features and achieve good detection results. In the 400 small nodule test data, the FROC curve is as shown:
It is worth noting that the algorithm takes 10 minutes to diagnose 200,000 lung nodules, which is far less than the time of doctor's manual diagnosis. It saves the doctor's time while improving the accuracy. It is really done in the diagnosis process. Doctor's assistant. The team also put the algorithm into the major hospitals in Shanghai for trial and use, embedded in the doctor's diagnosis process, and truly benefit patients.
Source: Heart of the Machine (WeChat almosthuman2014)
Fire Ball,Fire Fighting Ball ,Ball For Fire Fighting,Elide Fire First Aid Fire
NINGBO TOMAN IMP. & EXP. CO., LTD , https://www.tdotmfiresolution.com