How Shanghai Jiao Tong University Artificial Intelligence Laboratory uses AI to locate lung nodules

Cancer, like a devil in the dark, brings fear and despair to people. Lung cancer, as the type of cancer with the highest incidence and mortality rate in my country, hurts countless families. Nearly 600,000 people die of lung cancer in my country every year. However, the mortality rate of cancer is closely related to the period when cancer is first discovered. Early lung nodule screening can save countless people from pain and torture. Professor Xu Yi, Professor Ni Bingbing, Professor Yang Xiaokang, and Zhu Yumeng from the Artificial Intelligence Laboratory of Shanghai Jiaotong University cooperated with Dian Nei Technology to build an automatic positioning screening system for lung nodules using deep learning. It can effectively detect various types of nodules such as micronodules and ground glass in lung CT images, and reduce the occurrence of false positive misdiagnosis, realizing the desire of "early detection, early diagnosis, early treatment, and early recovery". The algorithm won the first place in the Tianchi Big Data Competition. This competition attracted more than 2,000 participating teams across the country, with a total prize pool of up to one million, and was open to major hospitals, universities, companies, and research institutes in the whole society.

The team used convolutional neural network technology in the field of computer vision to solve the problem of lung nodule detection and made innovations on multiple levels. 1) Combined object detection and segmentation algorithms to extract candidate nodules and generate a pool of candidate nodules with high recall rate. 2) Used a false positive attenuation network and adopted a multi-scale ensemble learning network model to improve detection accuracy and reduce the false positive ratio. 3) In terms of data processing, a generative adversarial network was used for data augmentation to improve the effectiveness of training.

The algorithm framework is shown in the figure

Data preprocessing

Through geometric transformations such as rotation and translation, data diversity is augmented for a limited number of positive samples, similar to how doctors analyze nodule areas from different perspectives and contexts. Generative adversarial networks (GANs) are used to generate new nodule positive samples from random noise and learn to generate nodule samples with new morphologies, thus deeply augmenting data diversity and improving model generalization capabilities.

Nodule Pre-detection

The 3D-Unet network structure is established. The main function of this segmentation network is to extract suspected candidate nodules, maximize sensitivity, and reduce missed detection rate. The network inputs three-dimensional data features and can be "observed" from multiple Z-axis dimensions, just like doctors observe nodules in multiple planes, so as to fully learn the difference between normal and abnormal textures inside the lungs and capture the diversity of nodules. For example, the density of ground-glass nodules is slightly higher than the surrounding area and is cloud-like, and the solid density of pure solid nodules is higher, similar to a single separated egg yolk.

Nodular sperm detection

Nodule precision detection uses three models to predict the probability of candidate nodules respectively, and gives the final probability according to the weight ratio between the models. The main advantage is that the negative samples undergo a learning process from easy to difficult, and the segmentation network and the subsequent false positive attenuation network complement each other. The model ensemble of multiple structural types, and the good performance of a single network, is similar to the process of multiple doctors reading the film independently, and the comprehensive reading results are given.

result

This algorithm won a great victory in the Tianchi Medical AI Competition jointly organized by Alibaba Cloud and Intel, standing out from more than 2,000 strong teams with a score of 0.732 and ranking first in the most important semi-final of the competition.

The model trained by the algorithm can better handle the features of nodules of different shapes and achieve good detection results. On the test data of 400 small nodules, the FROC curve is shown in the figure:

It is worth noting that the algorithm takes 10 minutes to diagnose 200,000 lung nodule films, which is much less than the time it takes for doctors to diagnose manually. While improving the accuracy, it saves doctors' time and truly serves as a doctor's assistant in the diagnosis process. The team has also put the algorithm into practice in major hospitals in Shanghai, embedding it into the doctor's diagnosis process to truly benefit patients.