Accrad - AI Radiology

Problem

The COVID-19 virus has spread to a global pandemic. Under a worst-case scenario with no interventions, Africa could see 1.2 billion infections and 3.3 million deaths, according to a report by the UN Economic Commission for Africa. Early recognition is vital not only for prompt treatment, but also for patient isolation and effective public health containment and response. 

Solution

We propose the use of AI-based CT and X-Ray image analysis for the recognition of COVID-19 infection in Africa. We have developed a deep learning algorithm to concurrently detect COVID-19 and 14 clinically important diseases in chest radiographs, at levels comparable to practising radiologists. 

Positive Impact

Our solution brings diagnostic expertise in Africa where radiologists are scarce. It drastically reduces diagnostic times, it reduces unnecessary COVID-19 tests, it improves the quality of diagnostic outcomes, and reduces the workload of radiologists. Hence, radiologists can diagnose more people with the same input. 

According to the WHO, 2/3 of the world population or 5.2 billion people have no access to radiology services. In Africa, 0 to 5 Computer Tomography Units are installed per 1,000,000 population. Radiography equipments are highly expensive. In addition, the lack of diagnostics expertise has created a crisis in Africa.

Chest radiograph interpretation is critical for the detection of acute thoracic diseases, including tuberculosis, COVID-19 and lung cancer, which affect millions of people worldwide each year. This time-consuming task typically requires expert radiologists to read the images, leading to fatigue-based diagnostic errors. The clinical workload of radiologists continues to increase globally and in Africa particularly. 

The lack of radiology expertise has left the African continent vulnerable to many complex diseases and outbreaks have repeated frequently over the last decades. COVID-19 has shown that diseases can develop to a global pandemic very quickly. We can't afford to leave the African continent alone on this fight. Our solution will dramatically increase diagnostic capacities and expertise across Africa saving lives and stopping disease spreads. 

We developed a deep learning algorithm which we call CheXRad to interpret Chest X-Ray and CT images. It can take as input any Chest X-Ray image and tell you in this image, what are the diseases that are present and what is my algorithm's predicted probability for each of those diseases. Some examples are COVID-19 or mass and nodule, which are cancerous, or enlargement of the heart and pneumonia. It concurrently diagnoses 14 different pathologies.

The algorithm is also able to convey some insight into its decision-making process by highlighting, for each disease, what particular part of the image it's looking at to make its decision. If I were a doctor with a suspicion of a certain disease, I could get a confirmation from the algorithm. Or if I were a clinician who were using this software, the algorithm could direct my attention to something I may have earlier missed.

CheXRad uses neural networks and datasets from African COVID-19 cases for disease detection and diagnosis. It can detect multiple radiologic findings, including lung consolidation, which indicates possible COVID-19 infected pneumonia. The algorithm uses 15 different machine learning models and can interpret 4.7 X-Rays per second, over 410.000 per day.

We are focusing on the African continent. Our AI-diagnostic offers radiologists a supportive tool during case overload, which can likely lead to low reading quality. It works with portable X-Ray and CT devices to fast-track test results in advance to definitive tests (e.g. PCR tests). It is of great support when regular monitoring is required among patients showing mild symptoms in order to identify and/or triage patients by the progression of symptoms. Furthermore, we provide the following benefits to radiologists:

• Drastically reducing disease diagnostic times,
• Predicting diseases at levels comparable to practising radiologists, 
• Bring diagnostic expertise in areas where radiologists are scarce,
• Reduce unnecessary tests,
• Improve the quality of diagnostic outcomes, and
• Reduce the workload of radiologists.

One of our co-founders is a radiologist helping us to understand their needs. Further, we are partnering with local radiology practises in South Africa. 

The average time for radiologists to complete labelling of 420 chest radiographs is normally distributed in a range of 180–300 minutes. Our method labelled the same 420 chest radiographs in 1.5 minutes. This is 120x - 200x times faster than practising radiologists. 

In Africa, 1,000,000 people share 0-5 Computer Tomography Units. People can't afford the high costs associated with radiology services. They travel long distances for a simple CT scan, sometimes 3 days of a bus drive. On the other side, there is a dramatic scarcity of radiologists on the continent. Radiologists are permanently overloaded with clinical and administrational work and can't serve their communities properly. Hence, diseases like Tuberculosis or COVID-19 remain undetected and the population suffers. With our solution, we help radiologists scale their expertise and diagnose more cases. 

The use of AI in radiology is quite new. There are a few companies worldwide working with similar algorithms, including us. 

What sets us apart is the method of applying (i) datasets for confirmed COVID-19 cases, (ii) the Machine Learning Frameworks in order to build a COVID-19 recognition model using multiple layers of Deep Neural Networks (the industry standard is 121 layers, we use far more layers, applying for a patent) based on input images from X-Ray scans and (iii) the hardware accelerators we use to build a COVID-19 recognition model and train it with full access to the latest CPUs, GPUs, and FPGAs, using the new programming language, Data Parallel C++ (DPC++). 

Working on the COVID-19 detection problem, we also use various hyper parameters to improve the performance of the deep learning models (we use 15 different models), focusing on the lungs. Specifically, we discovered how to precisely detect the lung location in the chest X-Ray, and crop out irrelevant areas by using optimised frameworks. 

We think the sum of these four features makes our solution belong to the top 5 solutions worldwide and top 1 solution on the African continent. 

We use a convolutional neural network-based method for recognition of COVID-19 in Chest X-Ray and Computed Tomography (CT) radiographs, and a method for medical image processing of large datasets related to COVID-19. The medical image processing method comprises of: 

1. Data Collection, 
2. Data Processing, and
3. Training a convolutional neural network.

See previous answer for more details. 

We are the prize winner of the Intel Innovation Competition Award with our technology. Here is a link that explains and provides evidence about our technology: 

Seeking Early Detection Using Medical Image Processing

We have also published an earlier version of our code on GitHub.

GitHub link

Our goal is to bring early disease detection to all places across Africa where radiologists are scarce. In Africa, less than 5 CT scanners are installed per 1,000,000 population. That has created a crisis and an extreme workload for radiologists. We can assist radiologists to reduce diagnosis times by over 90%, service more patients and avoid fatigue-based errors in fighting diseases like lung cancer, Covid-19, Tuberculosis and more.

We are in Beta phase. We are currently looking for research partnerships with healthcare providers and radiologists who are interested in working with us to validate the technology and prepare for the clinical trial.

12 months goals: 

1. Train the model with 2 million X-Ray and CT scans (we have currently scanned 117.000 images) to improve the algorithm.
2. Finish Clinical Trial
3. Start using the algorithm with 20 radiologists 
4. Apply for the patent

5 years goals:

1. The ultimate vision for us is to have this algorithm in the hands of any clinician across the African continent who would like to use it. 

12 months barriers:

1. We need funding to hire more technical staff
2. We need more confirmed COVID-19 CT scans from the African continent to train the model in an African context

1. We are applying for different programmes (like yours)
2. We are raising capital
3. We are partnering with a South African pathology laboratory to help us to get the clinical trial
4. We want to establish networks with universities and other organisations who can help us to validate our technology 

We are 4 founders (two technical, one radiologist and one business) at the moment and are interviewing talented AI engineers as well as Radiologists to join our team to move faster. We have bootstrapped 100% of our work until today.

We have all necessary skills inside our team. Our technical founders are serial entrepreneurs ans Intel Innovation Award winners, experts in Artificial Intelligence and Machine Learning. We have an Oncology Radiologists in our team and our CEO is a serial entrepreneur, a former SoftBank backed startup CEO and an ex-VC with strong consulting, banking and M&A experience.

• Intel: We use Intel CPUs, Cloud Infrastructure and more. We have won the Intel Innovation Award recently. See link above.
• Nvidia: We use Nvidia CPUs, Cloud Infrastructure and more
• Microsoft: We use Azure

For radiologists we save time significantly. We reduce costs for the healthcare system by avoiding diagnostic errors and radiologist overload. We speed up the diagnostic time by a factor of more than 100x and can help radiologists utilise their equipment much better. 

Our business model will comprise packages for CT scan analysis. Each package contains of 20 image scans for free and a fee for each image scanned afterwards. The costs will be significantly lower than the costs for radiologists and the healthcare system as of today.

In the beginning we will need several funding round to operate profyitabely and extend our growth.

Our business model will comprise packages for CT scan analysis. Each package contains of 20 image scans for free and a fee for each image scanned afterwards. The costs will be significantly lower than the costs for radiologists and the healthcare system as of today. 

We hope SOLVE can help us to get our Clinical Trial much faster with its knowledge and network. The prize money will also help us to buy more supercomputer cloud architecture as using those machines for our algorithm can quickly drive costs. We also hope to leverage SOLVE's network within the academia to validate our technology with experts.