- About
- Our solutions
- Resource Center
April 29 2022
|
Reading time 9 minutes
Contents
What do Physicians and Artificial Intelligence Face in Medical Image Analysis?
How do AI and Machine Learning Help in Medical Visualization?
In Which Medical Spheres the Neural Networks are Most Useful
Machine learning-based artificial intelligence (AI) can identify patterns in data and make decisions on their basis. It is trained on data templates without explicit programming and can improve as new information arrives1.
Medical visualization has to do with big flows of data, with which a physician works at various stages of examination. Here, artificial intelligence is used to accelerate the process and eliminate clinical uncertainty, which may arise when making diagnosis.
Medical image interpretation poses challenges both for a physician and for artificial intelligence.
What kind of difficulties can the physician encounter?
experts may have different diagnoses for the same scan;
a number of illnesses show minor changes at the early stages, which can be missed by visual examination;
a physician needs to look through more scans, often comparing them with previous examinations;
part of the time is spent on preparing reports.
What problems can artificial intelligence face?
Archives of medical images, which are the datasets, are unclear to the computer initially. They require preliminary processing. Medical experts manually mark each scan, checking it for pathologies and specifying diagnosis. The process can take long and depends on the diagnostician2.
AI can show high accuracy on a limited number of data but make mistakes in case of new samples. Developers are struggling to make the algorithm able to generalize and to be efficient in various clinical conditions3.
The dissimilarity of data is due to the peculiarities of the diagnostic equipment. The make and model of the scanner used in examination can influence the quality of scans. The heterogeneity of data can also be explained by the experience of the operator, different scanning parameters and diversity of patient groups (gender, age, disease)4.
The analysis of medical images includes the following tasks3,5:
Classification. In simple classification, the artificial intelligence assigns one of the categories to the diagnostic scan: norm or pathology. It is necessary to distinguish differences, for example, in the case of tumor neoplasms, a multiclass classification is used.
Localization It helps to determine the position of anatomic structures on the scan. For this, the sliding window method can be used: the neural network ‘slides’ over the image, using the window content as input data. The AI assesses whether the searched class is present in the given area. The algorithm then returns the location of the object: either as a central pixel or as a bounding box of four coordinates.
Segmentation. The algorithm ‘looks through’ each pixel of the image and assesses whether it belongs to the searched object.
The artificial intelligence raises the efficiency of the radiologist’s work due to routine process automation6:
areas of interest identification, including pathological parts;
medical image processing in the case of low quality of the scan: denoising, scan contrast and clarity enhancement;
the quantitative assessment of anatomic formation;
integration of the data obtained into radiology reports.
Medical image processing algorithms are becoming an important tool in the hands of the diagnostician. They are being included in clinical decision support system (CDSS). Using CDSS helps medical experts save time for solving non-standard diagnostic tasks.
Scientists from South Korea have developed a method for increasing the accuracy of breast cancer screening based on machine learning. The obtained results were published in The Lancet Digital Health. The research sample included 170,230 mammograms collected from institutions in South Korea, the United States, and the United Kingdom7.
Each mammogram was assessed by 14 radiologists for the probability of the malignant tumor presence in the breast. The neural network showed high diagnostic performance: the AUC indicator, which assesses the classification quality, was 0.9407.
The research has proofed that AI can be a support tool for a physician. Radiologists, who use machine learning methods, can identify breast cancer more accurately7.
Artificial intelligence does not know fatigue: it can analyze data continuously and make decisions in the conditions of increased load. It has become increasingly relevant during pandemic, when hospitals encountered a large flow of patients. Physicians have to identify persons with COVID-19 among those with respiratory diseases in order to provide medical assistance to them first.
Scientists from the US have used machine learning for sorting patients. The automated system takes into account clinical data: the number of lymphocytes, blood oxygen saturation level and presence of cardiovascular diseases. Chest X-rays were additional data source9.
Radiography can be applied in the conditions of limited resources, when medical institutions face medical staff and diagnostic equipment shortage. This method is characterized by relatively high speed and low radiation dose8.
The AI has conducted a complex assessment in 80 radiographies within 24 hours. The diagnostic accuracy was 95%: the algorithm detected patients with COVID-19 in 76 cases8.
Scientists from the Netherlands and China have developed an automated system based on the artificial intelligence to identify pulmonary tuberculosis. Methods and results of the research were presented in the European Radiology journal9:
to teach the neural network, radiologists marked manually the abnormal slices with pulmonary damage in tuberculosis;
to visualize the severity of tuberculosis, the researchers used the three-dimensional approach to reconstruct images;
the AI detected suspicious contagious areas and created colored attention maps;
the analysis of the CT data was carried out for 526 participants with clinically diagnosed tuberculosis.
Classification accuracy of the neural network for three independent datasets was from 81.08 to 91.05%. The authors point out that the developed system will help a physician to correctly assess tuberculosis severity and choose therapy9.
Cardiovascular disease visualization. Neural networks help to identify signs of acute pathologies, which require immediate help, such as myocardial infarction. Using machine learning, physicians assess hemodynamic changes on the scans, for example, coronary artery patency, heart valves pathologies or calcification. Visualization markers help stratify risks correctly and reduce mortality rate from heart diseases due to timely therapy20.
Visualization of lungs. A radiologist diagnoses respiratory disease by characteristic changes in lung tissue on a CT scan. The AI improves differential diagnosis by classifying pathological areas by density and texture characteristics. The analysis is performed at the pixel level, invisible to the human eye. The neural network simplifies visual quantitative assessments by calculating the volume of a suspicious focus22.
Mammography. The AI analyzes digital tomography image of a breast and detects suspicious areas among normal tissue. The neural network assesses the abnormality and classifies it as benign or malignant12.
SberMedAI solutions cover relevant areas of medical visualization. The services are used to diagnose widespread diseases:
CT Stroke: The AI marks a head CT scan, highlighting the areas of disturbed blood circulation in acute stroke;
CT of Lungs: the neural network identifies the percentage of lung damage in viral pneumonia, including COVID-19, and detects minimal nodules in lung tissue;
Mammography: the assessment of neoplasms in breast in accordance with the international BI-RADS scale and with the detection of the focus contours.
CT scans analysis with the use of AI helps to raise the security of radiographic examinations and obtain qualitative medical scan We will tell you more about the possibilities of artificial intelligence in computed tomography in another article.
The further development of artificial intelligence in medical visualization is aimed at solving the following problems13:
The development of useful databases. The quality of the classification by an algorithm depends on the training dataset. The storages need to contain organized and annotated images, which help the AI to learn to recognize a disease correctly and react to it.
Investment attraction. The analysis of medical images requires the engagement of experts at all stages, from marking data and manual localization to the assessment of the algorithm performance. Software products require technical support and necessary equipment.
Safety Compliance. In case of using the clinical databases, patients’ scans should be anonymized.
Complex approach to diagnosis. Apart from visualization data, the AI can take into account clinical indicators taken from an electronic health record. The integration of other sources of information is aimed at interpreting medical images.
The application of neural networks and machine learning in medical image analysis is in the state of active development. AI implementation in healthcare system is possible due to cooperation between researches, software developers and members of professional community.
Even now there are the first results: early and accurate diagnosis of a number of diseases, which means preserving health of a patient.
Sources
MachineLearning What it is and Why it Matters [Online]: SAS. URL: https://www.sas.com/ru_ru/insights/analytics/machine-learning.html. (accessed 17/03/2022)
Hyun Jin Yoony, Young Jin Jeong, Hyun Kang et al.
Medical Image Analysis Using Artificial Intelligence // Prog Med Phys. 2019; 30(2): 49-58.doi: 10.14316/pmp.2019.30.2.49.
Joseph R England and Phillip M Cheng. Artificial Intelligence for Medical Image Analysis: A Guide for Authors and Reviewers // American Journal of Roentgenology. 2019 212:3, 513-519. doi: 10.2214/AJR.18.20490
How to Help Radiology AI Cope with Different Scanners and Settings? [Online]: Quantib. URL: https://www.quantib.com/blog/how-to-help-radiology-ai-cope-with-different-scanners-and-settings. (accessed 17/03/2022)
Tova Linder and Ola Jigin, Organ Detection and Localization in Radiological Image Volumes. 2017.
Hans-Joachim Mentzel, Artificial Intelligence in Image Evaluation and Diagnosis // Monatsschr Kinderheilkd. 2 July 2021; 1-9. doi: 10.1007/s00112-021-01230-9.
Hyo-Eun Kim, PhD, Hyo-Eun Kim, Boo-Kyung Han, MD, et al. Changes in Cancer Detection and False-Positive Recall in Mammography Using Artificial Intelligence: a Retrospective, Multireader Study // Lancet Digit Health. 2020;2(3):e138-e148. doi: 10.1016/S2589-7500(20)30003-0.
Chris K. Kim, Ji Whae Choi, Harrison X. Bai, et al. An Automated COVID-19 Triage Pipeline Using Artificial Intelligence Based on Chest Radiographs and Clinical Data // npj Digit. Med. 2022:5:5. doi: 10.1038/s41746-021-00546-w.
Chenggong Yan, Lingfeng Wang, Jie Lin, Jun Xu, et al. A Fully Automatic Artificial Intelligence-Based CT Image Analysis System for Accurate Detection, Diagnosis, and Quantitative Severity Evaluation of Pulmonary Tuberculosis // Eur Radiol. 2022:32, 2188–2199. doi: 10.1007/s00330-021-08365-z.
L B van den Oever, M Vonder and M van Assen, Application of Artificial Intelligence in Cardiac CT: From Basics to Clinical Practice // Eur J Radiol. 2020 July;128:108969. doi: 10.1016/j.ejrad.2020.
F. Prayer, S Röhrich, J Pan, et al.
Artificial Intelligence in Lung Imaging // Radiologe. 2020;60:42–47. doi: 10.1007/s00117-019-00611-2
Krzysztof J Geras, Ritse M Mann and Linda Moy,
Artificial Intelligence for Mammography and Digital Breast Tomosynthesis: Current Concepts and Future Perspectives // Radiology. 2019;293(2):246-259. doi:10.1148/radiol.2019182627.
Geoff Currie, K Elizabeth Hawk, Eric Rohren, et al. Machine Learning and Deep Learning in Medical Imaging: Intelligent Imaging // J Med Imaging Radiat Sci. 2019;50(4):477-487. doi: 10.1016/j.jmir.2019.09.005.
