Cancer diagnosis through Artificial Intelligence

The ability to detect lesions as benign or malignant early on without biopsy is of utmost importance in cancer treatment. In a recent study, researchers used artificial intelligence and found promising results when programming an algorithm for breast cancer diagnostics.

A study uses an algorithm to support breast cancer diagnostics

Early diagnosis is of utmost importance for cancer treatment. For this reason, scientists are constantly working to develop better diagnostic methods. The ability to detect lesions as benign or malignant without biopsy would also be a decisive step forward. In a recent study, researchers used artificial intelligence and found promising results when programming an algorithm for breast cancer diagnostics.

A research team at the University of Southern California has set itself the task of shortening the time-consuming, multi-step process of elastography. To do this, they use an algorithm to obtain information from images more quickly.

The researchers wanted to determine whether the algorithm could be programmed to detect differences between malignant and benign lesions in breast scans. They used synthetic data instead of real scans.

The use of synthetic data

Professor Assad Oberai, main study author, cites the lack of availability of real data as the reason for using the artificial data: "In medical imaging, you can consider yourself lucky if 1,000 images are available. In such situations, where data are scarce, it is important to use synthetic data as well".

The researchers used over 12,000 artificial images to program their machine-learning algorithm, which they understand as a "convolutional neural network". Towards the end of the process, the algorithm was 100 percent correct for synthetic images.

In the next step, the researchers switched to real scans. Here, the research team had only 10 images, half of which showed malignant lesions and half of which showed benign lesions on one side. Professor Oberai says: "Here the accuracy rate was 80 percent. Next, we work on refining the algorithm by using more real scans".

The scientists firmly believe that they could have achieved a better accuracy rate if the algorithm had had access to more real data. The researchers also admit that their test scope was too small to predict the future capabilities of the system.

In recent years, there has been a growing interest in the use of AI in diagnostics. The authors of the study write: "Artificial intelligence is already being used successfully to evaluate images in radiology, pathology, and dermatology, often outperforming medical experts in terms of speed and accuracy".

The role of algorithms in medical imaging is becoming increasingly important

Nevertheless, Professor Oberai does not believe that artificial intelligence can ever replace trained medical professionals: "The general consensus is that such algorithms play a significant role, especially for medical imaging experts, on whom AI will have the greatest influence. Nevertheless, the algorithms will always be most useful if they are not used as a black box. What these algorithms have seen and how they came to their final conclusion must be understandable at all times in order to work with them".

The scientists hope that they will soon be able to diagnose other types of cancer with their new method. In their view, it should be possible to record changes in tissue and train the algorithm to detect them.

One limitation, however, is that due to the different interactions of cancer with its environment, an algorithm must overcome problems individually for each type of cancer. Professor Oberai is currently working on CT scans of kidney cancer in order to find out whether artificial intelligence could support diagnostics in this area.

Circumventing the solution of inverse problems in mechanics through deep learning: Application to elasticity imaging.
Dhruv Patel, Raghav Tibrewala, Adriana Vega, Li Dong, Nicholas Hugenberg, Assad A.Oberai
Computer Methods in Applied Mechanics and Engineering
Volume 353, 15 August 2019, Pages 448-466