Introduction
Step | Definition | How radiologists can contribute |
---|---|---|
1. Defining use case and the conceptual design | Defining clear, clinically relevant use case to achieving specific outcomes (e.g., increasing the speed, accuracy, and efficiency) and translating it into the conceptual design of the solution | Guiding the developers towards clinically relevant use cases (what problem to focus on) and how their solutions can potentially be used by radiologists [6] |
2. Data sourcing and curation | Collecting, selecting, cleaning, and organizing the data that is needed for the training and validation of the algorithm | Sharing their data (images and scans), thus being the connection between the available data in the medical world and the AI vendor [8] |
3. Labeling and establishing the ground truth | Defining the ground truth and (in case of supervised learning) labeling data | |
4. Training the algorithm | Configuring the algorithm (e.g., setting the parameters) and training it | [No specific role of radiologists is currently suggested in the literature] |
5. Testing and validating the AI application | Using appropriate and dedicated (reference) datasets to validate trained algorithms and ensure their accuracy and generalizability to clinical cases |
Methods
Company | Foundation year | Location | Company size (no. of full-time employees) | Company background | AI application (under development) | Radiologists involvement |
---|---|---|---|---|---|---|
Large multinational with diverse medical products | 1847 | New Jersey, USA | 50.000 | Large company that was originally founded from a medical background and currently tries to enhance their already existing physical medical products by implementing AI | AI application detects lung nodules on CT scans Algorithm: Deep Learning | Employment: both full-time and part-time contracts Status: workers, researchers No. of radiologists involved: ~1000 radiologists |
Large multinational from China | 2014 | Beijing, China | 300 | Company was created from a medical background from which they aim to enable doctors with higher efficiency and better diagnosis for patients through medical artificial intelligence | The application detects pneumonia cases in an accurate and timely manner, lesion quantifications, detects tiny lung nodules, tuberculosis from chest radiograph, multiple functioning for X-ray image detection, accurately detects bleeding area from CT stroke images and CT bone scans Algorithm: Deep Learning | Employment: part-time contracts and freelance assignments Status: researchers, consultants No. of radiologists involved: around 15 (an entire department is allocated to the radiologists) |
Medium-sized established AI vendor from Canada | 2007 | Calgary, Canada | 144 | A medical perspective has been the drive and background of this company, which enables radiologists to complete effective and precise analysis through artificial intelligence | The application assesses coronary artery disease using cardiac CT and detects cardiac abnormalities through MRI Algorithm: Deep Learning | Employment: freelance assignments Status: medical consultants No. of radiologists involved: radiologists only work for the company on a project basis. Projects could involve around 10–15 radiologists |
Medium-sized startup from California, USA | 2013 | CA, USA | 60 | Company founded from a technological perspective. Company tries to provide every care provider with the skills from highly trained radiologists through technology | A point-of-care ultrasound system that helps healthcare providers, even those with no experience, to conduct quick and accurate ultrasound exams at point of care of the cardiac functioning Algorithm: Deep Learning | Employment: freelance assignments with the company and external advisors Status: medical consultants, researchers No. of radiologists involved: two medical consultants within their team; several radiologists depending on the size of the project |
Small established AI vendor from Israel | 2006 | Haifa, Israel | 45 | Company builds vendor-independent images enhancements for diagnostic imaging. Company is built from a technological perspective | The application “enables the use of fast MRI protocols on MRI scanners of any vendor and model, by substantially increasing SNR and image quality” Algorithm: Machine Learning | Employment: radiologists are not part of the permanent team within the company; are only asked to give advice on certain concepts and products. Status: medical consultants No. of radiologists involved: depending on the project and kind of idea or product that are asked to be involved with |
Small Dutch Startup | 2012 | Rotterdam, The Netherlands | 35 | Medically centered company, which tries to use artificial intelligence to empower the radiologists by delivering fast, objective, accurate, and insightful medical reports | Detects prostate cancers from MRI images Algorithm: Deep Learning | Employment: both part-time and full-time contracts Status: medical consultants, researchers No. of radiologists involved: two full-time and around 12 part-time employees |
Small, established AI vendor in the Netherlands | 2014 | Nijmegen, The Netherlands | 32 | Company with a medical perspective from which it tries to use artificial intelligence to detect cancers earlier and thus improve breast cancer survival rates | Breast cancer detection and diagnosis for 2d and 3d mammography Algorithm: Deep Learning | Employment: part-time contract Status: medical consultants, researchers No. of radiologists involved: around 7 part-time employees, depending on the ongoing projects |
Small Startup in Lithuania | 2017 | Vilnius, Lithuania | 11 | Company started from a technological background from which it tries to implement this to create easy-to-use medical products for everyday clinical practice | A fully automatic computer-aided diagnosis (CAD) chest X-ray solution. It identifies chest X-ray images with no abnormality and produces preliminary reports Algorithm: Deep Learning | Employment: part-time contracts. Radiologists as one co-founder Status: medical consultants, researchers, and co-founder No. of radiologists involved: around 7 part-time employees, depending on the ongoing projects. Co-founder as part of the permanent team. |
Results
Role | Description | How and when the role happened |
---|---|---|
Step 1. Defining use case and the conceptual design | ||
Problem finder | Radiologists share their knowledge and ideas to define clinical problems, needs, and use cases; but they are not in charge of making the final decisions | • By hands-on radiologists, experienced in medical practice • Often as external experts (mainly to give their expert opinion) • Often as limited, occasional consultations (limited interaction / collaboration with the development team). |
Problem shaper | Radiologists are consulted to advice and give feedback on the medical problem and how the developers approach it; but they are not involved in making final decisions | • Predominantly present in companies with a medical background • Through long-term relations with radiologists (often as part-time affiliates and sometimes in-house radiologists) • Involves systematic, continuous interactions with the development team |
Problem dominator | Radiologists collaborate in defining and shaping the problem and make decisions regarding the definition of the clinical problem and the requirements of the application | • Present in large companies with extensive resources and staff • Through formal (full-time) position in the company • Formal responsibility in the development team • Often with extensive knowledge of AI and its applications |
Step 2. Data sourcing and curation | ||
Data champion | Radiologists identify relevant data and use their professional relations to get the data needed for the development | • Present in large companies with extensive resources and staff • Often when companies have strong medical background • In other cases, happened limitedly through casual consultations |
Step 3. Labeling and establishing the ground truth | ||
Data labeler | Radiologists annotate and label the data for the purpose of training the algorithm | • A common role across companies • Challenge of finding and paying for experienced radiologists • Sometimes using radiologists for the entire labeling and sometimes using alternative approaches such as using semi- or unsupervised algorithms, using automated tools for labeling, and training non-radiologists to do the labeling |
Data quality controller | Radiologists check the quality and process of labeling data (which is done by computational systems or trained employees) | • Present in companies with extensive resources and development projects • Often as in-house experts, involved in the development • Effective for companies with large-scale and extensive development projects |
Step 4. Training the algorithm | ||
Algorithm shaper | Radiologists give feedback on the understandability and accuracy of the outcomes of the algorithm for adjusting the training of the algorithm (e.g., selecting training datasets or tweaking parameters) | • Often through casual consultations with external experts • Sometimes as an iterative participation (not later at the end of the training) |
Step 5. Validating the AI application | ||
Algorithm tester | Radiologists test and validate the outcomes of the AI solution on its performance in various (complex) cases and under medical conditions. | • Present in almost all the cases • Often through in-house or affiliated radiologists who have a close collaboration with the vendors • In both clinical and non-clinical settings • Sometimes via “trial” versions offered to radiologists as potential users (to test it on their local data) |
AI researcher | Radiologists conduct (scientific) research on the performance of the AI solutions to produce scientific evidence and legal documents (e.g., for approval procedure) | • Through in-house (only in one company), but often via external collaborations (conducting joint research) • Often formalized around the approval process (e.g., FDA, CE marked) • Involves the publication of scientific papers • Active in cases with a strong medical background or connection with academic/research institutes |