Pulmonology is a field of medicine that deals with respiratory tract diseases, and the medical imaging used by pulmonologists is predominantly chest imaging: CXR, CT, MRI, PET, V/Q scanning, ultrasound, and the like. High-quality medical image analysis is crucial in pulmonary diagnostics and treatment. While the most conventional method of assessing lung tissue and surrounding structures is computed tomography, other types are used for additional insights and to accommodate individual contraindications.
What Steps Does Computer-Based Medical Image Analysis Include?
1. Improving image quality. When acquiring a medical image, radiologists have to strike the balance between quality and the permissible degree of exposure for the patient. This is especially critical when a person has to get multiple chest scans in a short period of time. Radiologists apply various techniques, such as decreasing the tube current and exposure time, but such manipulations often lead to increased image noise or appearance of image artifacts. In addition, object deformation during the scanning (for example the patient breathing) negatively affects image quality.
This is why quality improvement is a necessary step in medical imaging, allowing to reduce noise, remove artifacts, increase contrast, and prepare the image for segmentation.
2. Segmentation. At this stage, the acquired image is segmented into the lungs, lobes, and pulmonary segments, with lung tissue isolated from other tissue. Segmentation and pulmonary vascular tree modeling are important for clinical decision-making, providing the foundation for diagnostics and helping to course correct the treatment of lung diseases.
For computer-based segmentation, the difficulty lies in the wide variety of lung pathologies, including the ground-glass opacity, consolidation, cavity, nodules, pleural effusion, honeycomb, etc. The possible abnormalities differ greatly from normal lung tissue, manifesting as different shapes, textures, and attenuation in chest scans.
3. Quantification. The quantification of pulmonary imaging data refers to a number of tasks, from all-encompassing image fusion to analysis of lung segments to defining lung tissue form, size, texture, morphology, and dynamics. To this day, chest scans present special difficulties for imaging solution engineers due to the diversity of normal and abnormal lung tissue and the differing proportions of air, water, blood, and other fractions within the lungs depending on the lung disease.
What AI and machine learning bring to pulmonary imaging
Using cognitive computing in lung disease screening and diagnosis is a relatively recent development, but the technology is progressing fast and showing very encouraging results. AI-based 2D and 3D image analysis can match or even exceed the performance of trained radiologists and help them effectively detect lung abnormalities, such as early signs of lung cancer, emphysema, and pneumonia.
Machine learning in pulmonary image segmentation. Traditionally, the segmentation of chest scans was performed manually and required an excellent understanding of diagnostic criteria and a deep knowledge of anatomy. The advances in computer vision and machine learning made it possible to speed up or even automate the process of lung segmentation.
Machine learning algorithms are trained on rich sets of imaging data that contains labels for healthy lung tissue, various abnormalities and pathologies, and surrounding tissue. As the algorithm learns to recognize patterns in labeled data, it also learns to assign correct anatomic classes to elements in new, unlabeled chest scans.
The machine learning-based segmentation method involves analyzing every pixel (in 2D images) or voxel (in 3D images) of the image, which requires significant computational power. However, with modern computers constantly increasing their processing capabilities and the costs of such calculations falling, the highly accurate segmentation that machine learning solutions provide has become accessible to more healthcare providers.
Deep learning in image classification, detection, and segmentation. Deep learning has emerged as a powerful tool for advanced medical image analysis, namely the tasks of image classification, segmentation, and abnormality detection. In this approach, the algorithm learns patterns and feature representations from processing raw, unlabeled data — thousands upon thousands of chest scans.
As a recent example, we can look at the usage of deep learning in assessing CT images of lungs infected with COVID-19. Multiple CT scans throughout diagnosis and treatment have become the clinical norm for handling patients with the novel coronavirus. This practice puts a lot of pressure on radiologists where image segmentation and quantification are still completed manually.
A group of scientists from the department of radiology in the Shanghai Public Health Clinical Center suggested a deep learning-based approach to computerized segmentation and quantification of such images. The system detects and quantifies infected regions and their volumetric ratios in relation to the lung. A comparison of automatic versus manual image segmentation was carried out on 300 CT scans of COVID-19 patients, delivering the average Dice similarity coefficient of 91.6 percent.
AI in pulmonary image analysis. The global pandemic has undoubtedly created an additional incentive for performance-driven innovation in pulmonary imaging. An automated tool for quick and reliable detection of pneumonia — a common symptom of coronavirus —would help doctors to triage incoming patients more efficiently and provide life-saving treatment to those who need it.
Various companies have started looking into enhancing lung image analysis solutions with AI. Amazon Web Services teamed up with researchers at University of California to study the capabilities of artificial intelligence in spotting COVID-19 pneumonia on chest X-rays . The resulting algorithm was trained using 22,000 notations by human radiologists, managing to correctly indicate areas of pneumonia, even in images acquired at different locations with varying technique, contrast, and resolution.
Myrian, a solution for visualization and analysis of medical images, applies AI and deep learning to enable a wide array of features that help streamline the workflow of COVID-19 responders, including automatic lung segmentation, automatic calculation of global lung volume and ling reserve ratio, visualization of healthy and pathological lung areas, lung density histograms, and more.
The role of machine learning, deep learning, and artificial intelligence in pulmonary radiology is increasing. As recently as January 2020, before the pandemic took hold, one-third of surveyed hospitals and imaging centers were reporting that they use these technologies to drive patient care or business operations.
The great clinical potential of automated image analysis lies in the growing accuracy and efficiency of the algorithms, which lead to accurate assessment of the patient’s state and a more personalized approach to care.
There’s still space for growth, however. The current approaches are highly fragmented and often designed with a specific lung disease, failing to address other abnormalities that may be in the image. The lack of a unified platform with an approved user interface limits the coverage of computer-assisted analysis and doesn’t let it become a standard method in clinical practice.
Thus, future efforts in automated pulmonary image analysis will likely focus on improving the quality of analysis in parallel with increasing efficiency and interoperability through better integration with other hospital systems.
This article was written by Irina Demianchuk, for Oxagile, a custom software development company. For more information, visit here .