Bench Talk

(Source: zapp2photo - stock.adobe.com)

Artificial intelligence (AI) is a growing trend around the world. A number of high-tech sectors are adopting AI with the aim of benefitting the industry, whether it be time, money, or accuracy benefits. AI is becoming an integral part of the chemical and pharmaceutical sciences. On the one hand, you have artificial neural networks offering companies ways of designing and synthesizing new drugs. On the other hand, machine learning algorithms offer a way to detect cancer cells in a patient with a much higher degree of accuracy. We’re here to talk about the latter.

Scientists have used many ways of testing and looking at cancer over the years, but microscopy and other imaging techniques have been some of the key methods. The days are gone when a scientist would look under the microscope and come to a conclusion without any computer assistance. In recent years, a number of computer software programs enabled scientists to look at the shape, size, and morphology of imaging samples, including cells. However, many of these programs still require human input to characterize where the points of interest, in this case, the cancer cells, start and where they stop. These imaging methods have been designed for a range of analysis equipment, from the simple lab bench microscope to MRI scanners.

So, while there are effective methods for cancer diagnosis—which is why humans as a whole have gotten better at recognizing and understanding the disease—most of these methods are still prone to human error. Even a small error can cause a misdiagnosis. Machine learning emerged in the last few years as a potential solution to this. Results, to date, have shown the ability to analyze imaging samples and pinpoint the presence of cancer cells with a high degree of accuracy.

Several chemical sensors can detect whether a patient has cancer. While clinicians can check for specific biomarkers in patients’ blood, biological samples are inherently complex. While tests can detect specific biomarkers, biological samples are inherently complex, and the analysis of human biological fluids can sometimes produce error-prone results. So, clinicians can employ machine learning algorithms in conjunction with early warning chemical tests to cut through the ‘noise’ of the tests and analyze the data points of interest for determining whether a patient has cancer.

Cancer cells exhibit certain characteristics that distinguish themselves from healthy cells. These characteristics are often a way of determining whether a patient has cancer, alongside specific biomarkers present in the blood when a patient has the disease. Especially from an imaging perspective, the physical features of both healthy and cancerous cells are an easier way of physically seeing whether a patient has cancer or not.

For example, where normal healthy cells of the same type tend to have the same shape and size—often spherical/oval in nature unless they are specialized cells—cancer cells tend to have a much different (i.e., more random) shape and size, which can stick out around healthier cells. Additionally, in healthy cellular systems, the division of cells tends to be controlled, and the arrangement of the cells is organized. On the other hand, cancer cells divide at much higher rates and tend to be very disorganized.

Another feature that cancer cells possess is that they tend to have large, variable-shaped nuclei, whereas healthy cells only have a small, regular shape nucleus. In addition, there tends to be a loss of features within cancer cells—which is why they are dangerous, as this loss of features is why cancer cells cannot perform specific functions like healthy cells. All these differences and characteristics of both healthy and cancerous cells can be used, analyzed, and compared by machine learning algorithms providing that the software has enough data.

Machine learning algorithms offer a way to analyze cancer cells better and determine whether cancer cells are present in a patient. Machine learning algorithms work by taking the historical data and matching it with the data from the current analysis. The ability to compare the historical data with the new data enables the algorithms to detect whether the system is normal—in this case, healthy cells—or whether there are abnormalities—i.e., cancer cells.

To do this, the machine learning algorithms need to be fed data from previous studies, which includes the different sizes, shapes, and surface morphologies of both cancerous and healthy cells. By doing this, the algorithms can quickly and easily identify which cells within an image are healthy and which are potentially cancerous. By providing an accurate and statistical way of analyzing the cells, the algorithms mitigate human error when determining if cells are indeed cancerous or if further tests need to be performed to confirm if a person has cancer.

But it’s not just imaging methods where machine learning can help with a cancer diagnosis. In recent years, a number of early warning point-of-care devices have been created that can detect whether a patient has cancer much earlier on. Many of these devices are based around microfluidic systems, where the inside is coated/functionalized with specific surface receptors that will attach to any cancer cells. So, the receptors do need to be specific to the cancer being targeted, but these systems essentially act as a series of point-of-care nanosensors that can provide an early warning signal, enabling clinicians to treat the disease much earlier on, and in turn, increasing the chance of survival.

So, where does machine learning come in? Quite a bit of data can be collected from these platforms (and chemical tests in general). Trying to spot the trends between the different data sets to provide an accurate diagnosis is not the easiest task—as these range from the size and morphology of the cells to the gene expression and the extent of growth/division within cellular groups.

Point-of-care devices can be coupled with imaging methods to analyze the data from a chemical perspective and simultaneously image the sample while it’s being analyzed.. So, by segmenting the image into slices, alongside using machine learning algorithms, the principles mentioned above about deciphering the difference between health and cancer cells can also be applied to some point-of-care devices as well. So, there is the possibility in the future of combining both chemical and imaging diagnosis methods to create platforms that can provide quantitative and qualitative analysis.

Is machine learning the future for cancer diagnosis? It’s an open-ended question with an open-ended answer at this stage. A great deal of interest and work is going into machine learning and other AI algorithms, with interest across the medical and pharmaceutical sectors. Machine learning holds great promise in the future of cancer diagnosis as it is beneficial on both the chemical testing and the imaging side. The scope is significant, and machine learning might be more useful in one area than another at a clinical level.

The use of machine learning and AI algorithms, while growing in use, is still in its infancy. Although many sectors are starting to adopt it more, the medical community must scrutinize technologies more because of the potential issues surrounding misdiagnosis and patient welfare. Nevertheless, there is a drive to have more accurate analyses for cancer and other diseases. Machine learning offers the potential for this and the benefit of removing human bias.

Obviously, there is also the ethical side to consider with medical diagnoses, and AI approaches will probably still need human input from a trained clinician to confirm the results. Otherwise, issues could occur in the event of a misdiagnosis or an issue with the software, which can happen with any type of technology. While medical technology can fail, there is usually a human backup to rectify errors in most clinical settings. So while machine learning could provide all the analyses, it may be the case that human input is still required from an ethical standpoint.

If we can handle the ethical considerations and the algorithms are made to be accurate and reliable, then there should be no reason why we won’t see machine learning being employed in cancer diagnoses in the future in one form or another. But, only time will tell the extent to which medical professionals will adopt AI in oncology and wider clinical settings.