Computer vision is a field of artificial intelligence that trains computers to obtain high-level understanding from digital images or videos. It enables the automation of tasks that human visual systems can do. With the development of algorithms and the combination of other technologies, computer vision has been moving from entertainment fields to industry practices, creating more opportunities to empower companies in different fields.
The COVID-19 pandemic raises an alarming concern that the industry needs to reimagine healthcare in totality. It has made it clear that data is critical to make decisions promptly and remotely. The development of telemedicine and IoMT devices also provide more possibility for remote diagnosis. In this process, computer vision, supported with other AI techniques for processing and analyzing images, has played a role in providing disease prevention, improving diagnostic accuracies, and raising an alarm by monitoring patients’ conditions 24/7.
Diabetes is the leading cause of kidney failure, accounting for 44 percent of new cases. If diagnosed early, we can possibly slow down the progression of end-stage kidney disease. With help of computer vision, we can analyze Renal Ultrasound and MRI images to accurately predict the stages of Chronic Kidney Disease(CKD) progression. As an industry practice, computer vision powered software has been developed to control the spread of COVID19 as well:machine learning models are trained to recognize faces wearing masks with 95% accuracy; infrared thermography is applied to screen fevers among crowded people; if you don’t consider legal and ethical concerns, drones could also be applied to detect patients that have tested positive in public areas.
Neurological conditions affect our feeding behavior. dysphagia can lead to malnutrition, dehydration, and aspiration, and that can lead to aspiration pneumonia, which is the leading cause of death in patients. But the pain point of research on feeding behavior is the cumbersome tracking of markers manually. To solve this pain point, Kavi Global used videos of rodents feeding on kibble to characterize feeding behavior. Kavi developed a near-automated, deep learning-based solution for detecting and tracking fiducial markers in the videos, resulting in a more efficient and robust process, with potentially 300+ times reduction in data processing time. 100% accuracy has been achieved on manually tracked data.
Pain monitoring and differentiating pain from distress are important to decide the right treatment. Early-life pain has the potential for lasting effects across the neural axis and adverse neurodevelopmental consequences. Because of their under-developed pain inhibitory pathways, neonates are 30-50% more sensitive to pain than adults and have reduced pain tolerance compared to children. Inability to self-report pain makes neonates vulnerable to under-recognition, and under or overtreatment of pain. The current pain assessment is mainly done by nurses manually. Limited by nursing personnel, only 10% of neonates receive a daily assessment of prolonged neonatal pain. When trying to apply computer vision to PICU, there are two pain challenges:concerns on safety, ethnicity, and traceability and technical limitations on accurately classifying pain when neonates are lying on the hospital bed. Kavi Global is developing computer vision-based neonatal facial pain coding systems on fetus and neonates videos to find explainable solutions on pain identification and improve the accuracy of detection when neonates/fetuses have their faces partially covered by hands or other objects.
Example of computer vision in healthcare:detect abnormal X-Ray image
Data source:Kaggle dataset-National Library of Medicine, National Institutes of Health, Bethesda, MD, USA and Shenzhen No.3 People’s Hospital, Guangdong Medical College, Shenzhen, China
Monitoring and Preventative Maintenance
Railroad track inspection needs to be conducted frequently to identify track defects. The current methodology of railroad maintenance involves manual inspection by trained railroad personnel or slow hi-rail equipment, often resulting in a costly and time-consuming process. Kavi Global proposed the application of comparatively inexpensive cameras fitted in regular locomotives and took the video as input. With Kavi’s experience in forecasting and failure detection models, the health information of rail can be monitored and provides continuous improvement.
Defect Detection in Manufacturing Process
Defect detection in wafer fabrication processes enables the detection of malfunctioning procedures and thus drives proper intervention to solve the problem, to increase the fab yield.
To optimize yield, enhance wafer inspection and connect the semiconductor and electronics supply chain, computer vision techniques have been applied to identify defect wafer from qualified ones and classify the defect type. High accuracy rates have been achieved by applying machine learning models on the features extracted (out of the identified defects, 99.3% are actual defects after manual inspection, and 98.6% of defects in training samples have been successfully identified).
Example of computer vision in manufacturing: detect types of metal surface defects
Data source:Kaggle dataset-GC10-DET-Metallic-Surface-Defect-Datasets
About the Authors
Ziwen Qin is a Consultant, Business Analytics at Kavi Global. She has a Masters degree in Supply Chain Management from University of Michigan – Ross School of Business. She is proficient in optimization and data modeling and passionate about delivering value through emerging technologies.
Trishla Mishra is a Consultant, Business Analytics at Kavi Global. She has a Masters degree in Business Analytics from Oklahoma State University. She has been working in the field of analytics for 5+ years, primarily in data engineering, predictive analytics and deep learning.