AI software development process in medical imaging field

Companies can generally be divided into two categories, business-driven and technology-driven. The customer group of business-driven companies is the general public. For example, Ele.me. The purpose of technology is to better support business development, while theFor example, database software companies are technology-driven, and some companies are in between. For example, Alibaba, when the business develops to a certain level and the accumulated technology is quite strong, the company can expand the company through technology realization..

When the AI ​​tide hits and many industries want to be part of it, are AI software companies business-driven or technology-driven?Although AI software serves hospital institutions, it ultimately serves the general public. In addition, with the emergence of AI frameworks, the AI ​​threshold is not so high, so I prefer it to be business-driven.

Make a request:

Since it is a business driver, we should understand what users need. You can refer to my previous article [Application of AI in the Whole-Process Work Cycle of Imaging Equipment]. This article mainly focuses on how to proceed with the next development when the demand is determined.

Musk has said that asking questions is more important than the ability to solve known problems!The perfect application of AI in the field of medical impact requires deep insights into the field.

Select tool:

The next step is to choose an AI development framework. "Framework" allows us to focus on doing business without caring about technical details. "Doing business" is to convert business into processes and translate processes into codes. Behind the AIis mathematics, including mathematical operations, optimization algorithms, probability theory, etc.

The current AI frameworks are crowded with stars, Tensorflow, Pytorch, MXNet, Caffe, Baidu's Feipao, Huawei's MIndScope, and MegEngine's MegEngine. These frameworks are all open source. In order to attract user groups, some frameworks provideFull-process development documents, AI courses and free computing power.

The choice of framework depends on the technical accumulation of the team. Some frameworks have a very high degree of integration, and training can be started with a simple command line. Such frameworks are suitable for teams with little previous AI technology experience. In addition, the model community supportsDegree is also an important factor. With high community activity, some important models can be reproduced and open sourced soon.

At present, Tensorflow is the most used in industry, and Pytorch is preferred in academia. Domestic Feipu provides AI computing power, platforms, frameworks, tool kits, and training courses, which are comprehensive and similar in style to Pytorch, suitable for beginners.

AI development is divided into two relatively independent processes: training and inference. Training does not need to be integrated into the product, so you can choose a development language that is inconsistent with the product development language. Currently, Python is the most used.Provided for different language inference frameworks, the result of AI training is a model file and a parameter file. Although the output file formats of different frameworks are different, there are tools to convert them into a common format: ONNX, so companies often use itOne kind of framework is used for training, and another framework is used for reasoning. Because the reasoning process is relatively simple, some companies have accumulated strong technology, and even write their own reasoning framework based on CuDNN.

Result analysis:

When we are excited to complete an AI product, reality hits us hard, the effect is not satisfactory, overfitting or underfitting?We started to adjust parameters wildly and change the model. If the effect is still not satisfactory, we need to change the direction of thinking. Is there a problem with the data?

Data is very important for AI. Data scarcity is a problem faced by many companies. However, medical data has its particularity. It is difficult to generalize the pre-trained model of large-scale data sets to the medical field. The current effective method isUse data enhancement methods to augment data, such as using basic image transformation, or using GAN to generate data. The amount of data in the early stage is small, and ensemble learning can be used to use a combination of multiple simple models; with the increase of data, iterativeIncremental way to retrain and increase model complexity.

Model complexity and inference speed are a double-edged sword. Model complexity is high, its parameters are large, and there are many floating-point operations, resulting in slow inference speed. However, high model complexity also means high model accuracy. Therefore, it is necessary toWhen choosing a technology, make a choice based on your own business.