Swift Medical partnered with Sama to cost-effectively scale annotations for their wound monitoring app without compromising on clinician-level accuracy.
The field of healthcare boasts some of the most impactful — at times even life-saving — applications of AI. Universities, healthcare incumbents, and new entrants in the field are turning to technology to innovate, with significant benefits for patients and healthcare providers alike.
For patients, AI can provide increased access to information, quicker diagnosis, and second opinions for more accessible and affordable care. For healthcare providers, AI-powered medical devices provide better outcomes and enhanced safety in surgical procedures and the treatment of diseases.
Among the most promising clinical applications of AI is diagnostic imaging: using computer vision to help diagnose, treat, and monitor diseases such as brain disorders, liver problems, and cancer cells. For patients suffering from injuries, diagnostic imaging also has the potential to help them monitor and receive care for their healing wounds.
The challenge of clinical variability at scale
The field of wound care is ripe with potential for AI, given its shortage of specialists (1 for every 500 patients) and minimal training for general practitioners (by some estimates, less than nine hours of formal wound training).
This labor shortage is compounded by the inconsistency and variability of clinical decisions. Even among expert wound care physicians, determining the presence and regions of various tissue types can vary from one expert to the next — made worse by manual and error-prone methods of assessment such as measuring wounds with paper rulers.
Ultimately, imprecise measurement and human subjectivity can compromise clinical decisions, leading to improper treatment, adverse events, and poor patient outcomes.
For Swift Medical, computer vision has the potential to help solve the massive challenge of clinical variability at scale. Their mobile app uses artificial intelligence to help patients and clinicians monitor the healing of wounds. Clinicians take a picture of a wound with their smartphone, and Swift provides accurate measurements, progress tracking, and other relevant information to evaluate wound health.
In order to provide wound measurement and diagnostics that rivaled or even surpassed the expertise of trained physicians, Swift Medical needed datasets with clinician-level accuracy.
Striking a balance between label cost and quality
Swift Medical initially turned to in-house experts to label their datasets. And though internal clinical teams undoubtedly had the expertise to turn around high-quality annotations to help their models behave predictably, when it came time to scale the data, they needed an approach that would scale with them.
Pivoting to an external labeling provider was not working for them either, as Swift’s Senior Machine Learning Researcher Dhanesh Ramachandram explained:
“We used other annotation providers in the past, but our experience with Sama was superior due to the direct communication we had with the labeling team, who were with us throughout the project.”
In-house data labeling had proven to be expensive, but Swift Medical knew that poorly labeled data is just as costly, with the potential to negatively impact patient and healing outcomes. Swift Medical needed a partner who could strike this balance of providing clinician-level accuracy for their labels, affordably, and at scale… so they reached out to Sama.
How Sama helped Swift Medical scale annotations without compromising quality
Sama was engaged to help Swift Medical scale the anonymized data used to train their AutoTissue AI model, which determines the tissue types present within a wound. This model needed a dataset with images segmented by different regions around the wound area, and proper labeling for the wound type.
In order to ensure clinician-level quality, Sama introduced rigorous training, communication, and quality assurance cycles:
- Specialized, in-depth annotator training: Swift’s clinical team provided bespoke training and detailed instructions on how to label tissue types to ensure that annotators were subject matter experts. To start, the Sama team was tasked with labeling 30% of the dataset; a subset to validate that they were up to speed and skilled enough to proceed with the full dataset.
- Continuous feedback loops: Sama annotators remained in constant communication with Swift via Slack, email, and calls. This enabled fast adoption and quick alignment on edge cases so annotators could iterate on instructions. Over time, continuous feedback decreased time per task and increased task complexity in the workflow. In Ramachandram’s words:
“Sama gave us visibility into the data labeling process, with tight QA feedback loops to ensure the high standard of quality we required for our models.”
- Rigorous quality control: Sama implemented a multi-step quality process, with AutoQA features complemented by Quality Control experts from both Sama and Swift reviewing samples and providing feedback for annotators to improve. Swift was more hands-on for the first ~1,000 images, with progressively less support as Sama became more skilled at delivering high-quality labels.
These QA and feedback cycles were key in delivering a high-quality dataset to Swift medical, which ultimately boasted an average internal quality score of 97%. In addition to this, the project saw a steady week-over-week increase in throughput with fewer rejections: from week 1 to week 10, throughput increased by 230% while rejected tasks decreased by 66.7%.
High-quality data labels weren’t the only prerequisite for Swift’s dataset — they wanted to ensure that their model would work well for individuals with all skin tones. Swift Chief Engineer and Co-Founder Justin Allport explained:
“Sama’s ethics and mission-driven nature appealed to us. We needed an AI model that performed well for all our customers’ patients, regardless of their skin tone. Working with an ethical organization to help us develop an ethical model was a priority.”
Ethical development plays a role in every aspect of AI at Swift, from data sourcing, to data labeling, to how the model’s output is used in the field. Sama’s ethical labor practices were an important part of that equation for Swift.
What high-quality labeled data can bring to healthcare
Expert training, constant communication, and rigorous quality assurance combined with an ML-powered platform enabled the Sama team to deliver a dataset with the high degree of precision that Swift needed: clinician-level accuracy for highly nuanced workflows at scale.
This dataset now powers the Swift mobile app, which autonomously and accurately identifies wound tissue types and proportions.
This accurate information is empowering clinicians to make data-backed treatment decisions and create more personalized care plans, improving the overall quality of care they can deliver to their patients.
What’s more, patients gain a better understanding of their wounds which leads to better patient outcomes, satisfaction, and engagement in their own care — a win-win scenario that decreases litigation risks for hospitals, skilled nursing facilities, and home health organizations to help avoid costs.
Read the executive summary for the case study below.