The authors’ suggestion of using fluorodeoxyglucose-positron emission tomography (PET) in the future for prognosis and monitoring is wonderful. We wish to add that the “rim sign”, a slight and continuous fluorodeoxyglucose uptake at the border of a peripheral lung consolidation, is easily recognizable on fluorodeoxyglucose PET/CT (though data on sensitivity/specificity are not available). When present, it strongly suggests pulmonary infarction and is observable even without suggestive finding of pulmonary infarction. The reverse halo sign would also be seen. Though highly sensitive, use of PET/CT for primary detection of COVID-19 is constrained by poor specificity as well as considerations of cost, radiation burden and prolonged exposure times for imaging staff. However, in patients who may require nuclear medicine studies for other clinical indications, PET imaging may yield the earliest detection of nascent infection in otherwise asymptomatic individuals. This may be extremely vital for immunocompromised patients, including those with coexistent malignancies, where the early diagnosis of infection and subsequent initiation of care needed will contribute vitally to improving outcomes and reducing morbidity and mortality.
Role of optical thermal imaging and other remote patient monitoring devices
Lukose et al stated that the currently popular method of collecting samples using the nasopharyngeal swab and subsequent detection of RNA using real-time PCR has false-positive results and a longer diagnostic time frame. Various optical techniques such as optical sensing, spectroscopy and imaging show great promise in virus detection, and the progress in the field of optical techniques for virus detection unambiguously show great promise in the development of rapid photonics-based devices for COVID-19 detection. They also provided a comprehensive review of the various photonics technologies employed for virus detection, especially the SARS-CoV family, such as near-infrared spectroscopy, fourier transform infrared spectroscopy, raman spectroscopy, fluorescence-based techniques, super-resolution microscopy and surface plasmon resonance-based detection.
Gomez-Gonzalez et al reported a proof of concept of optical imaging spectroscopy for rapid, primary screening of SARS-CoV-2. A study by Shah et al found that home pulse oximetry monitoring identified the need for hospitalization in initially non-severe COVID-19 patients when a cutoff SpO2 of 92% was used and that home SpO2 monitoring also reduced unnecessary emergency department revisits. McKay et al stated that due to its portability, affordability and potential to serve as a screening tool for a conventionally lab-based invasive test, the mobile phone capillaroscope could serve as an important point-of-care tool and that the simplicity and portability of their technique may enable the development of an effective non-invasive tool for white blood cell screening in point-of-care and global health settings. This would be extremely useful in the COVID-19 pandemic scenario as white blood cell monitoring forms an essential part of COVID-19 management and follow-up[41,42].
Infrared thermography has been considered a gold standard method for screening febrile individuals during pandemics since the SARS outbreak in 2003. Khaksari et al showed that in addition to an elevated body temperature a patient with COVID-19 will exhibit changes in other parameters such as oxygenation of tissues and cardiovascular and respiratory system functions. They also promulgated a compelling need to develop a new technique that would have the ability to screen all these signals and utilize the same for early detection of viral infections. In their opinion, keeping the advent of wireless technologies in mind, the development of such sensors that have point-of-care home-accessible capabilities will go a long way in better managing the increasing numbers of patients with COVID-19 who are opting for home quarantine and that this will eventually reduce the burden on the healthcare system.
The COVID-19 pandemic is changing the landscape of healthcare delivery worldwide. There is a discernible shift toward remote patient monitoring. It is pertinent to note that a large number of remote patient monitoring platforms are already utilizing optical technologies. This area of research has great potential for growth, and the biomedical optics community has great prospects in the development, testing and commodification of new wearable remote patient monitoring technologies to add to the available healthcare armamentarium and contribute to the rapidly changing healthcare and research environment, not just for the COVID-19 era but far beyond.
Various other ingenious methods/modalities have been used for early detection/screening for COVID-19. These include smartwatches, smart phones and other intelligent edge devices. Mishra et al developed a method utilizing data from smartwatches to detect the onset of COVID-19 infection in real-time that detected 67% of infection cases at or before symptom onset. They stated that their study provided a roadmap to a rapid and universal diagnostic method for the large-scale detection of respiratory viral infections in advance of symptoms, highlighting a useful approach for managing epidemics using digital tracking and health monitoring. Seshadri et al stated that when used in conjunction with predictive platforms, wearable device users could receive alerts when changes in their metrics match those related to COVID-19 and that such anonymous data localized to regions such as neighborhoods or zip codes could provide public health officials and researchers a valuable tool to track and mitigate the spread of the virus. Their manuscript describes clinically relevant physiological metrics that can be measured from commercial devices today and highlights their role in tracking the health, stability, and recovery of COVID-19 + individuals and front-line workers.
Schuller et al in their paper tilted ‘COVID-19 and Computer Audition: An Overview on What Speech & Sound Analysis Could Contribute in the SARS-CoV-2 Corona Crisis’ provided an overview on the potential for computer audition, i.e., the usage of speech and sound analysis by AI, to help in the COVID-19 pandemic scenario and concluded that computer audition appears ready for implementation of (pre-)diagnosis and monitoring tools and more generally provides rich and significant, yet so far untapped, potential in the fight against COVID-19 spread.
AI in COVID-19 imaging. Telemedicine has advanced by leaps and bounds. AI algorithms enable faster diagnosis (including remote diagnosis), with a fair degree of accuracy. While the application of AI to medical imaging of cancers and other diseases is being developed over the past decades, the recent COVID-19 pandemic hastened the: (1) Need; (2) Development; (3) Training; and (4) Testing of AI algorithms, within a relatively shorter time-span of less than 2 years. This was extremely beneficial for radiologists and other physicians involved in performing rapid diagnosis, keeping in mind this was a time when there was immense overloading of the healthcare system. The benefits including for management were obvious. However limitations such as: (1) Limited datasets; (2) Inaccurate execution of training and testing procedures; and (3) Use of incorrect performance criteria needed to be dealt with. The above limitations can be overcome by the utilization of federated learning[48,51,52].
The technique of federated learning was originally pioneered by Google as an application of their well-known MapReduce algorithm and allows for iteratively training a machine learning model across geographically separated hardware, including mobile devices. The machine learning algorithm is distributed, while data remains local. It can be employed for both statistical and deep learning. Despite its drawbacks, specifically wide-area network bandwidth limits computation speed, federated learning appears to be a great way forward, especially for multicenter collaborations, getting around the ‘tricky’ data privacy issue and enabling algorithms/outcomes with much more accuracy than otherwise possible.
If AI is to make an even greater impact, Merchant et al suggested getting down to the basics and incorporating time tested key medical ‘teaching’ and/or key ‘clinical’ parameters, including prognostic indicators, for more effective AI algorithms and their better clinical utility. They also stated that “Artificial Intelligence needs real Intelligence to guide it!”. Combining the wisdom gained over the years with the immense versatility of AI algorithms will maximize the accuracy and utility of AI applications in medical diagnosis and treatment modalities. We have gained wisdom regarding COVID-19 imaging over the past few years and should utilize the same for creation of better algorithms for screening/detection/prognostication and management.
El Naqa et al, as part of a Medical Imaging Data and Resource Center initiative, noted that the pandemic has led to the coupling of interdisciplinary experts that include: (1) Clinicians; (2) Medical physicists; (3) Imaging scientists; (4) Computer scientists; and (5) Informatics experts, all of whom are working towards solving the challenges of the COVID-19 pandemic, specifically AI methods applied to medical imaging. They stated that the lessons learned during the transitioning to AI in the medical imaging of COVID-19 can inform and enhance future AI applications, making the entire transition more than every discipline combined to respond to emergencies like the COVID-19 pandemic. AI has been used in multiple imaging fields for COVID-19 imaging.
The model by Manokaran et al could achieve an accuracy of 94.00% in detecting COVID-19 and an overall accuracy of 92.19%, which was based on DenseNet-201. The model can achieve an area under receiver operating characteristic curve of 0.99 for COVID-19, 0.97 for normal and 0.97 for pneumonia. Their automated diagnostic model yielded an accuracy of 94.00% in the initial screening of COVID-19 patients and an overall accuracy of 92.19% using chest X-ray images.
Kusakunniran et al proposed a solution to automatically classify COVID-19 cases in chest X-ray images using the ResNet-101 architecture, which was adopted as the main network with over 44 million parameters. A heatmap was constructed under the region of interest of the lung segment to visualize and emphasize signals of COVID-19. Their method achieved a sensitivity, specificity and accuracy of 97%, 98% and 98%, respectively. Rao et al stated that separable SVRNet and separable SVDNet models greatly reduced the number of parameters while improving the accuracy and increasing the operating speed.
Yi et al utilized a large CT database (1112 patients) provided by the China Consortium of Chest CT Image Investigation and investigated multiple solutions in detecting COVID-19 and distinguishing it from other common pneumonia and normal controls. They compared the performance of different models for complete and segmented CT slices, in particular studying the effects of CT-superimposition depths into volumes, on the performance of their models and showed that an optimal model could identify COVID-19 slices with 99.76% accuracy (99.96% recall, 99.35% precision and 99.65% F1-score).
Chaddad et al investigated the potential of deep transfer learning to predict COVID-19 infection using chest CT and X-ray images. They opined that combining chest CT and X-ray images with DarkNet architecture achieved the highest accuracy of 99.09% and area under receiver operating characteristic curve of 99.89% in classifying COVID-19 from non-COVID-19 and that their results confirmed the ability of deep convolutional neural networks with transfer learning to predict COVID-19 in both chest CT and X-ray images. They concluded that this approach could help radiologists improve the accuracy of their diagnosis and improve overall efficiency of COVID-19 management.
Cho et al performed quantitative CT analysis on chest CT images using supervised machine learning to measure regional ground glass opacities and inspiratory and expiratory image matching to measure regional air trapping in survivors of COVID-19. They summarized that quantitative analysis of expiratory chest CT images demonstrated that small airway disease with the presence of air trapping is a long-lasting sequelae of SARS-CoV-2 infection.
Fuhrman et al developed a cascaded transfer learning approach to extract quantitative features from thoracic CT sections using a fine-tuned VGG19 network where a CT-scan-level representation of thoracic characteristics and a support vector machine was trained to distinguish between patients who required steroid administration and those who did not. They demonstrated significant differences between patients who received steroids and those who did not and concluded that their ‘cascade deep learning method’ has great potential in clinical decision-making and for monitoring patient treatment.