By Abdulaziz Alanazi
Artificial intelligence has grown to unimaginable lengths in recent years. Digital AI models create almost life-like videos and images and solve complex problems and prompts, but a new application of artificial intelligence in medical diagnosis and treatment is being used for patients. AI can use patients’ information and tests such as MRI and CT scans to make an accurate conclusion and diagnosis of the patient, which may be more efficient and accurate than being diagnosed by a doctor. To make AI models that are capable of analyzing complex images and diseases, we use machine learning and deep learning, which is how AI can be made to be able to process information through pattern recognition. The increasing power of artificial intelligence holds much promise for the improvement of the medical industry and the health of patients.
CT and MRI scans create digital images of the patient’s body, which are used by doctors. CT images are 2D images that are made through a process called reconstruction, which combines X-ray imaging through different angles to create a cross-sectional image and capture the specific structure the doctors are looking for. Since CT images are 2D, the pixels in the image determine certain information, such as thickness, contrast, and details of the body part. CT scan vs. MRI: What’s the difference? And how do doctors choose? (2022). On the other hand, MRI uses a magnet to create a magnetic field, which causes atoms in your body to align in the same direction as the magnetic field. Then, the MRI machine sends radio waves, which cause the atoms to move out of position. The radio waves stop, and the atoms go back to their original positions, which return radio signals. The radio signals are received by the computer, which converts the signals into an image on the screen, which can then be examined using magnetic resonance imaging (MRI). (n.d.).
The image can be put into an artificial intelligence model, which can determine the patient’s diagnosis. There are two ways that AI can analyze the image. The first way uses machine learning, which needs a region of interest (ROI) to be determined from an outside source and then identifies features such as volume, shape, size, intensity, and location to conclude. Although usually CT and MRI images have multiple regions of interest, using machine learning models is less efficient than deep learning, which, unlike machine learning, doesn’t need a region of interest to be determined and can be given the entire image to analyze. It uses layers of classification, which is called the neural network, and each layer gets more specific until a clear conclusion is reached. Deep learning is especially good for CT and MRI scans because of RAGCN (Region Aggregation Graph Convolutional Network), which can segment the image into different sections and analyze each ROI separately without any human intervention. Although deep learning may be better for images with more than one ROI, both methods are viable and will increase work efficiency and accuracy in the field of radiology.
About 340,000 Americans die from misdiagnosis, and diagnostic errors from radiologists are a factor in those deaths. Heavy workloads and understaffed work environments make it difficult for radiologists to ensure an accurate diagnosis and give patients the best quality care. To counter this, early developments in AI are being used to diagnose diseases and cancers. For example, gastric cancer is the fifth most common cancer, with one million new cases and 780,000 deaths each year (Chen HY, 2022). The stage at which the cancer is diagnosed can greatly determine its survivability; however, it is difficult to get an early diagnosis because of how similar it is to inflammatory lesions. So, the use of AI has been implemented. Hirasawa (2018) created a deep learning model using a convolutional neural network (CNN). They trained the model with 13584 images, then took 2296 gastric cancer images from 69 patients to test the AI model. The AI model was able to correctly diagnose 77 gastric cancer tumors in 42 seconds. Another example is liver cancer. Nishida (2022) created three AI models, one with 24,675 images, one with 57,147 images, and one with 70,950 images of hepatocellular carcinoma (HCC), metastatic tumors, hemangiomas, and cysts, to train the three AI models to see their effectiveness in accurately identifying lesions in comparison to human physicians. The AI models 5 specialized physicians and 3 non-specialized physicians were given 55 video images of 55 different patients and expected to identify what type of lesion it is. The median accuracy between the four lesions in the AI models was 80%, 81.8%, and 89.1%, respectively. The five specialized physicians’ median accuracy was 67.3%, and the median accuracy between the three non-specialized physicians was 47.3%. The performance of the three AI models passed the test of human physicians and proved AI can help prevent human error in diagnosing patients (Nishida, 2022).
A limitation of the AI models was identifying rare types of tumors and lesions. AI models need lots of training images for the neural network, especially for the diagnosis of lesions with a lot of variation. So, rare tumors with few cases and images cause the models to not be able to accurately give a diagnosis. An example is intrahepatic cholangiocarcinoma (ICC), which is only about 3-5% of primary liver cancer, and with little training images, the AI model was able to accurately diagnose 71.5% of ICC tumors, which is lower than the AI model’s score in the other four lesions (Nishida, 2022). Artificial intelligence has come a long way, but there still needs to be more investment to make 100% accurate models in the future.
Although AI has shown great results in diagnosing patients, one might say that physicians and health workers may be against AI because they are worried artificial intelligence will take their jobs or the trouble of trying to work with artificial intelligence, but according to a survey conducted by Sarwar, S., Dent, A., and Faust (2019), the evidence disagrees. 487 physicians with different amounts of experience from all over the world took part in the survey, which asked various questions about artificial intelligence’s future in medicine. When asked about their concern about AI replacement, 184 responded that they felt AI would not have any effect on employment, 205 felt that it would increase the number of jobs and demand for workers, and only 95 of the 487 respondents had concerns. Instead of AI disrupting workflow, 346 respondents feel that AI will increase efficiency and accuracy when working. According to the results of the survey, many physicians believe that artificial intelligence will have a positive impact on the medical industry, such as creating new work opportunities and increasing productivity and quality of work.
In conclusion, using artificial intelligence systems with deep learning and machine learning has proven to be more effective and accurate than human physicians. Such as the Hirasawa (2018) model being able to analyze 2296 images in just 42 seconds. Many physicians believe that AI will create more job opportunities and increase job efficiency. However, there is a limitation to how we can train these models for rare disease types because of the lack of training images for the model, causing the model to achieve low accuracy in identifying these diseases, like intrahepatic cholangiocarcinoma (ICC). This imperfection in AI should not discourage its use; it rather means AI has a future worth investing in to create a near-perfect model to 100% accurately diagnose patients to prevent human error and more deaths caused by misdiagnosing.
References
- CT scan vs. MRI: What’s the Difference? And How Do Doctors Choose. (2022, December 8). Memorial Sloan Kettering Cancer Center. https://www.mskcc.org/news/ct-vs-mri-what-s-difference-and-how-do-doctors-choose-which-imaging-method-use#:~:text=CT%20scans%20take%20a%20fast,a%20CT%20scan%20cannot%20detect.
- Magnetic Resonance Imaging (MRI). (n.d.). National Institute of Biomedical Imaging and Bioengineering. https://www.nibib.nih.gov/science-education/science-topics/magnetic-resonance-imaging-mri#:~:text=MRIs%20employ%20powerful%20magnets%20which,pull%20of%20the%20magnetic%20field.
- Hosny A, Parmar C, Quackenbush J, Schwartz LH, Aerts HJWL. Artificial intelligence in radiology. Nat Rev Cancer. 2018 Aug;18(8):500-510.
- Yu, C., Helwig, E.J. The role of AI technology in prediction, diagnosis and treatment of colorectal cancer. Artif Intell Rev 55, 323–343 (2022).
- Hirasawa, T., Aoyama, K., Tanimoto, T. et al. Application of artificial intelligence using a convolutional neural network for detecting gastric cancer in endoscopic images. Gastric Cancer 21, 653–660 (2018)
- Nishida, N., Yamakawa, M., Shiina, T. et al. Artificial intelligence (AI) models for the ultrasonographic diagnosis of liver tumors and comparison of diagnostic accuracies between AI and human experts. J Gastroenterol 57, 309–321 (2022).
- Ghaffar Nia, N., Kaplanoglu, E. & Nasab, A. Evaluation of artificial intelligence techniques in disease diagnosis and prediction. Discov Artif Intell 3, 5 (2023).
- Chen HY, Ge P, Liu JY, Qu JL, Bao F, Xu CM, Chen HL, Shang D, Zhang GX. Artificial intelligence: Emerging player in the diagnosis and treatment of digestive disease. World J Gastroenterol. 2022 May 28;28(20):2152-2162.
- Sarwar, S., Dent, A., Faust, K. et al. Physician perspectives on integration of artificial intelligence into diagnostic pathology. npj Digit. Med. 2, 28 (2019).
