At DeepTek, we aim to leverage AI to democratize radiology. Our USP is a close collaboration of experts in machine learning, radiology, and technology. Radiologists have pain points in their day-to-day workflow, and they know best what kind of product they would use. Our technology and data science teams find a way to achieve the exact features that would be useful for radiologists.
DeepTek’s co-founders have unique skills in key domains of radiology, data science, business planning, and strategy. The leadership and their teams of radiologists, data science experts, technology, and finance ensure rapid outcomes across multiple divisions, from data curation, testing, and validation to strategic deployments.
DeepTek provides a unique offering. We call it the Radiology Workflow Optimization Platform. The platform has medical imaging AI, image and non-image related workflow AI, intelligent business tools, a smart analytics wall, and a smart communication command center with a chatbox, all bundled in a thin, agile cloud framework with multiple microservices components. The platform’s success is supported by a unique human-in-the-loop model, which ensures the AI keeps improving with use and feedback from radiologists.
Since inception, we have partnered with major technology companies across the globe. Our strategic partners include NTT DATA, NVIDIA, Docnet, Edison GE, and Nobori.
DeepTek’s end-to-end radiology workflow platforms can help diagnose 22+ pathologies from X-rays, MRI, and CT scans. From chest X-rays, we can screen COVID-19, TB, emphysema, hernia, fibrosis, atelectasis, etc. We have also developed a model to diagnose the different grades of knee osteoarthritis from knee X-rays.
These models form the machine learning core of our comprehensive radiology workflow optimization solution, Augmento, which aims to reduce the burden on radiologists and save time and resources. Augmento is a cloud-based platform, and it uses AI to pre-screen all studies and present their results to the radiologist, who can approve or modify them at the click of a button. A standardized, high-quality report is generated instantly and made available to the radiologist for a final sign-off. The AI organizes all the cases for reporting in a smart worklist that presents critical cases first, ensuring prompt service to those who need attention the most. The analytics dashboard provides a visual overview of the reported cases, patient demographics, imaging modalities, etc. It has an inbuilt notification system that supports sending push notifications to physicians, administrators, and patients.
Genki focuses on managing the workflow efficiently in population screening programs, where the volume of data is large and quick results are required. It can evaluate a large volume of images in a matter of seconds, making it a formidable tool for fighting large-scale diseases like COVID-19 and TB. We painstakingly designed our radiology and AI solution to manage epidemic and pandemic situations successfully. The solution can be deployed on mobile equipment, allowing screening of patients in hotspots at their doorstep. The platform’s smart notifications also remind patients of follow-up clinical exams after a predefined time interval. Together, these features support treatment planning and disease containment, which are crucial for governments and NGOs working to resolve public health crises.
The success of our AI models
Many challenges hinder the performance of models when they are deployed in the field in live programs. These include a lack of sufficient training data, noisy annotations, training images not matching those observed in practice (domain shift), etc. Additionally, chest X-rays often capture foreign objects like chest leads and pacemakers, which can throw off AI models trained to detect specific pathologies such as granuloma and cause them to give a higher number of false positives.
We took the time to address these deployment problems. Our models include a domain shift detector and a foreign object detector. We also used a deep learning tool called “attention” to increase our models’ sensitivity. This technique makes the algorithm pay more attention to the more influential areas of a chest X-ray image. The output the model gives is thus more refined. We also use distributed deep learning to alleviate privacy concerns of the medical data we use from various sources to build more robust AI models. Using techniques like federated learning and split learning, there is no need to share raw data and risk the privacy of patients.
We are currently working on a model for diagnosing scoliosis from X-rays and chest illnesses from CT scans. We also plan to deploy our solution on edge devices and expand our capacity to evaluate millions of X-rays.
We strongly believe that leveraging both human expertise and AI can truly augment diagnostic radiology. In imaging diagnostics, deep learning models are already at par with expert radiologists in the controlled laboratory test environments. However, most of these models fail to work just as well when deployed in real clinical settings because they fall prey to issues such as biases due to inadequate (limited) training data and lack of generalizability. Our solution incorporates feedback given by radiologists to improve the machine learning models continuously. We attribute our success in the practical implementation of our solutions to the expert-in-the-loop approach.
Deployment in public screening programs
We worked with the Clinton Health Access Initiative and the Chennai City Corporation to screen Chennai for TB under the Stop TB program. Our work is one of the first prospective studies that successfully deployed an AI model for screening TB on the field. Our models demonstrated that they save costs in running TB screening programs and also help save radiologists’ time and effort.
DeepTek also partnered with BHS to provide an AI-assisted instant triage facility for the screening of tuberculosis and other diseases from digital chest X-rays using its Genki screening solution in Rajasthan. Realizing the need for accessible, migrant-friendly, and low-cost primary health care services in underserved areas, BHS and AMRIT Clinics have been vocal to support the underserved communities of South Rajasthan. In these areas, public systems have limited reach, and the illness load is high.
We have also deployed our solutions in Malaysia, Indonesia, and Japan.
Presenting our research at the world stage
Our team participates in and contributes to major research conferences in radiology and computer science around the globe, enabling us to incorporate the most recent advances in these fields into our offerings. Our team’s strong academic foundations have numerous research presentations, publications, and academic credentials to their credit. Some of the conferences we have participated in this year include the Union World Conference on Lung Health, RSNA Virtual Conference, NVIDIA GTC 2020.
Our research publications are listed here: https://www.deeptek.ai/research.