Using Artificial Intelligence to Identify Cancer

New and more advanced technologies such as artificial intelligence (AI) are being incorporated to improve clinical diagnostics, giving scientists and doctors an opportunity to evolve healthcare. Recently, researchers have developed an AI algorithm in order to reduce the chances of a false diagnosis, and to help identify cancer with more ease.

How do these algorithms work?

As like how doctors and nurses must be educated in order to do their jobs, AI algorithms must be competent enough in order to mundane tasks such as image analysis and decision making. In order to do so, we must instruct the AI explicitly what they should identify in an image given to the algorithm. 

In order to produce a successful AI algorithm, the computer system must be inputted with data that has a label or annotation and data points that the algorithm can recognize. Once the algorithm identifies and are exposed to enough parameters, the response of the algorithm is analysed to ensure that they are analyzing the issue accurately. This is an example of a typical algorithm “exam” used to test the competency of the program. Usually the test data inputted have answers known to the programmers. This allows them to adjust and modify the algorithm based on the response it produces.

Deep Learning based Automatic Detection (DLAD)

One example of these algorithms is Deep Learning based Automatic Detection (DLAD). This was developed by researchers at Seoul National University Hospital and College of Medicine in late 2018, in order to analyze chest radiographs and detect abnormal cell growth (ie- potential cancers). When the algorithm’s performance was compared to multiple physician’s detection abilities on the same images, it outperformed 17 out of 18 doctors (Greenfield, 2019). Figure 1 demonstrates how minuscule the difference is between normal cells and potentially dangerous cells. The left image is the data inputted into the algorithm, the right image is the region of potentially dangerous cells identified by DLAD.

Figure 1: Image taken by Sean Wilson, Harvard graduate

Figure 1: Image taken by Sean Wilson, Harvard graduate

Even with the best doctors, these details are hard to differentiate accurately due to human error. Using AI algorithms can help combat this difficulty in order to identify possible cancerous cells at earlier stages, in order for the cancer to be treated better and more effectively.

Lymph Node Assistant (LYNA)

Another such example is the Lymph Node Assistant (LYNA) algorithm, created by researchers at Google AI Healthcare in late 2018. This algorithm is programmed to analyse the histology slides of stained tissue samples in order to identify metastatic breast cancer tumors based on lymph node biopsies. Interestingly, this algorithm was able to identify suspected regions of metastatic breast cancer that are indistinguishable to the human eye based on the biopsy samples. When LYNA was tested on two datasets, it was shown to accurately identify a sample as cancerous or noncancerous correctly 99% of the time (Greenfield, 2019). Furthermore, LYNA was able to half the average slide review time when LYNA was used to help analyse stained tissue samples.

In the short run, these such algorithms have shown their capability to increase physician accuracy. They can be used by doctors to double-check their diagnoses and to interpret patient data quicker without losing accuracy. In the long run, these algorithms once approved by the government, could function on its own in a clinic, allowing doctors to focus their attention on cases that these algorithms are unable to solve. 

These two algorithms shows us a glimpse of the potential strengths of algorithms in medicine, however as this is a relatively new development in the field of AI, there are some difficulties implementing them. One reason is that regulating these algorithms is a complicated process. The U.S food and Drug Administration (FDA) as approved some assistive algorithms but currently, there are no universal approval guidelines that exist. Furthermore, as the programmers that develop these algorithms are not always doctors, in some cases these computationalists may need to learn more about medicine in order to develop an algorithm more suited for doctors. 

Thumbnail credit: