Where is the AI stuck in reading pictures and seeing a doctor?
arts/
In recent years, as the core force leading the transformation of the new generation of industries, artificial intelligence has also demonstrated new application methods in medical care, and new forms of business have been born in deep integration. Especially in the aspect of impact recognition, the application of artificial intelligence technology to medical imaging diagnosis is the most widely used scene in the medical field.
In the early stage of the development and application of artificial intelligence medical imaging, pulmonary nodules and fundus screening were popular areas. In the past two years, with the continuous maturity and iteration of technology, major AI medical imaging companies are also expanding their business radius. Breast cancer, stroke And bone age testing around bone joints has also become a key area for market players.
However, while the application of artificial intelligence to impact recognition may sound like a fiery scenario, the reality is disappointing. In fact, medical imaging AI in hospitals at all levels is not uncommon these days. However, without exception, they all entered the hospital in the name of "trial". To some extent, this reflects the current status of the implementation of medical imaging AI - there is more than free trial, but not enough money to buy .
The reason is that, on the one hand, as an emerging product, AI products still have a lot of immaturity. For example, according to the latest version of the Procurement Law in 2014, the bid cancellation conditions stipulated that if there are less than three suppliers who meet the professional requirements or suppliers who respond substantially to the bidding documents, the bid will be cancelled. This means that, among the few competitors for medical AI products, if there is one who "falls off the chain", the bidding process may start all over again.
For example, the AI-assisted diagnosis project of Chongqing Daping Hospital announced its purchase intention on November 24, 2021, and the pre-bid results were not announced until March 10 this year. Among them, there have been two rejections, because the company's technical support materials and financial audit reports have not passed the qualification review, so that there are less than three remaining suppliers. The project purchased the most mature CT pulmonary nodule products in medical imaging AI. The above-mentioned sales executive told Jiemian News that a similar difficulty in recruiting and purchasing is especially true in the case of coronary and head and neck products with higher technical thresholds and fewer manufacturers.
On the other hand, in terms of the "nutrient" of AI products, there are still problems such as small data volume, few dimensions, low quality, and "data silos". There are also gaps in the informatization level of hospitals at all levels. According to data from Ouyi Think Tank, in 2019, the penetration rates of medical image transmission and archiving systems installed in China's tertiary, secondary and first-level hospitals were 87.8%, 62.2%, and 40.1%, respectively. In fact, until today, in the real world, due to the decentralization of data, low replication costs, and value aggregation, data is still highly dispersed, and "data silos" are still obvious.
Finally, medical care is a matter of life. False negatives of AI medical images are obviously very important. Even if there is 1% of missed diagnosis, it may cause huge harm. In addition, even if there is only 1% of missed diagnosis, doctors still need to review all the films again. . Therefore, only zero false negatives can really help doctors save time and effort.
As a new generation of infrastructure construction, the application of artificial intelligence in the medical industry will bring changes to the operation of traditional medical institutions and effectively relieve the pressure on medical resources in the long run . But at present, artificial intelligence may still have a long way to go in order to achieve autonomous film reading.