Since the outbreak, a number of technology companies have joined the fight against the epidemic of battle.
Among them, the epidemic investigation is the top priority of the battle, while CT imaging has become an important basis for the new crown pneumonia, screening and treatment of the disease.
However, in the current critical period of diagnosis and treatment of the epidemic, a huge stock of new patients and the overall number of patients, doctors need multiple CT imaging of patients with various advanced follow-up compared to disease progression and therapeutic effects on patients for accurate assessment.
Using the traditional visual means of viewing medical imaging, doctors not only a huge amount of work, it is difficult to achieve accurate and timely comparative assessment of the patient's condition. In the whole society to fight the epidemic of medical resource constraints, the case of doctors work overload, excess CT imaging of the front-line fight against SARS work created a huge demand for medical resources challenge.
In fact, AI carried out by medical image analysis and diagnosis, has a lot of application cases. Just yesterday (2.28), Baidu jointly with heart health, officially launched the "pneumonia screening and disease based on CT images of the pre-assessment AI system", and has been put into use in the Affiliated Hospital of Chenzhou, Hunan Shonan Institute.
A, from seconds to minutes
According to reports, a new crown pneumonia patient CT images in about 300, which put huge pressure on physicians in clinical diagnosis, doctors CT image of the naked eye a case analysis takes about 5-15 minutes.
The first landing in the Affiliated Hospital of Hunan Institute of pneumonia this AI screening and pre-assessment system, to be completed lesion detection of CT images of patients in tens of seconds, the outline of the lesion, lung density distribution histogram and lung lesions calculation and display a full quantitative index number, volume, accounting and other lung.
Wherein the lesion detection system precision and recall on the test data set 92% and 97%, respectively, so that to prevent undetected in ensuring high precision on the basis of lesion detection.
In addition to rapid detection and identification pneumonia lesions, to provide for the diagnosis of disease lesion number, size, proportion and other pulmonary quantitative evaluation information. The system is also supplemented by histograms and density distribution of lesions in the lungs visual display means superimposed outline for pre-screening and clinicians provide patients with the diagnosis of pneumonia disease qualitative and quantitative basis, to enhance the efficiency of medical diagnosis and evaluation.
此外，该系统采用的深度学习算法模型充分训练了所收集到的高分辨率和低分辨率的 CT 影像数据，能极好地适应不同等级 CT 影像设备采集的检查数据，有望为医疗资源受限和医疗水平有限的基层医院提供有效的肺炎辅助预诊断工具。
二、用开源对抗封闭
很多公司也想打造自己的肺炎 CT 影响分析模型，然而从头训练的成本较高，不能够及时发挥作用。
为此，百度和连心医疗采取了开放的态度，在业内首次开源上述系统中的肺炎 CT 影像分析 AI 模型——肺炎 CT 影像分析模型（Pneumonia-CT-LKM-PP）。
不仅如此，其预训练模型也已在百度 EasyDL 上开放，开发者可通过在 EasyDL 图像分割模型中，选择“肺炎 CT 影像识别专用算法”，少量数据训练即可获得基于实际场景进一步优化的模型。
对于想亲自上手「肺炎 CT 影像分析模型（Pneumonia-CT-LKM-PP）」的开发者，百度也给出了详细的教程。
1、定义待预测数据
如果没有自己的数据，也可以用百度提供的 demo.dcm 医学影像练手。
2、加载预训练模型
百度的 PaddleHub 提供了病灶分析和肺部分割的 Module，即 Pneumonia_CT_LKM_PP，包含病灶分割和肺部分割 2 个模块，都是基于 UNet 进行一系列优化。
3、预测
PaddleHub 对于支持一键预测的 module，可以调用 module 的相应预测 API，完成预测功能。
4、 后处理
通过一定的后处理，将肺部分割结果映射到原图上，再将病灶分割和肺部分割融合到一张图上可视化。
代码传送门：
三、AI 战疫，为爱而战
随着临床诊断数据的积累，新冠肺炎的影像学大数据特征逐渐清晰，相信 AI 在肺炎筛查领域发挥的作用会越来越大。
据百度介绍，该系统后续还将陆续于湖北、成都等地医院部署，其在线版本也将对全国定点收治医院免费开放，有利于医疗人员基于该系统开展远程会诊协作，提高基层医院的病情诊断和救治能力，进而有望降低患者在转诊、巡诊等过程中产生的交叉传染风险。
也期待更多的医院和算法研究者参与到基于 AI 的医学影像大数据抗疫产品研发中来，为抗疫临床研究和临床产品研发贡献力量。
众志成城，打赢疫情阻击战。
