Author: Arya : Medical Image AI (Ph.D.) is studying, and the research direction is unsupervised deep learning; multi-modal; PET-CT. I write this article to sort out my learning ideas, and I hope that I can give a quick guide to friends who have just started to enter this industry .
From the perspective of scientific research, the traditional application of deep learning in medical image segmentation, such as segmentation and registration, has been weakening, and it is basically combined with U-Net. MICCAI is the top conference in the field of medical imaging. Judging from MICCAI 2019, most papers still use various modified versions of U-Net/FCN, or add a GAN. The more innovative work appears in horizontal applications such as tumor prediction/metastasis prediction/multi-modal. The trend of MICCAI 2019 is consistent with the message I received at China's first medical image conference in Shanghai this year:
1) Prediction is a new research direction;
2) Multimodal and multi-tumor detection (rather than primary tumor detection) is more in line with industry needs
It is worth noting that medical image processing has always been a field closely related to academia and industry. Computer-aided diagnosis Computer Aided Diagnosis Systems is a concept that existed when imaging equipment was first available. Instead of using traditional methods + prior information for analysis, now we have more deep learning methods. As for how to transfer from learning to production, the first medical imaging conference also mentioned many policy supports that are being revised or introduced.
From the perspective of employment, domestic companies such as United Imaging, SenseTime, and Tencent are relatively large companies in the field of medical imaging. Among them, United Imaging is a subsidiary company of United Imaging (one of the largest imaging equipment companies in China), led by Boss Shen, one of the most powerful people in the field of medical imaging, and has huge resource advantages; SenseTime is a hot artificial intelligence company; Tencent doesn't need to say much. In addition, there are many start-up companies in China with a financing of around 100 million or less, such as LinkingMedical and Vision Technology. Vertically subdivided products, such as Lianxin's radiotherapy target area delineation, and visual pulmonary nodule detection. Basically all companies are still in the stage of researching and developing products, looking for landing scenarios, and waiting to obtain clinical qualifications. All companies are trying to figure out ways to: 1) Pull hospitals to discuss cooperation (get desensitized data for model training); 2) Dock imaging equipment end-to-end cooperation (based on automatic diagnosis as a function key of imaging equipment). At present, the industry lacks a unified standard. A more direct idea is to make an industry standard in a certain imaging field, such as PET-CT/PET-MRI, combined with its monopoly advantages in imaging data.
Thanks to the big cows who led the team to promote it, I personally feel that doctors have welcomed this technology very much, and doctors and computer scientists are working hard to overcome the problem. To give a small example, I went to a tertiary hospital in a third-tier city to do minor surgery at the beginning of the year. Before the operation, I chatted with the anesthesiologist. When I talked about doing medical AI by myself, they were very excited, saying that this must be the trend in the future. My mother is also a practitioner in the medical industry. Over 50 years old, she can still talk to me about the application of AI in imaging. She is also very interested in this technology.
Personally, research in this field is different from natural images and computer vision. In addition to mastering basic research methods, it is more important to: 1) data with accurate annotations 2) clinical supports.
For example, in our laboratory, the idea to be done must be approved by a clinical professor. If he thinks it has clinical value, he will provide us with desensitized data, and we will do it, and the paper will not have clinical value. It will be accepted and the most important basis for whether the product can be produced; at the same time, it will also provide a lot of suggestions for the clinical professor in the research process. In my opinion, this is often more useful than the comments given by my computer boss. .
Therefore, the advice to friends who want to enter this scientific research industry is:
1. Enter a company or laboratory with medical resources
2. Focus on clinical value
3. Innovate from imaging and clinical perspectives, combine the advantages of deep learning methods, abandon alchemy.