Improving Medical Image Analysis Accuracy From 30 to 90 percent for a US Diagnostics Provider

About the Client

The Client is a diagnostic company that provides remote diagnostic support to healthcare organizations via a distributed network of medical specialists. The Client’s digital platform allows partnering clinicians to submit clinical imaging data and receive timely expert feedback, speeding up life-saving medical decisions.

Remote Diagnostic Platform Defects Slowed Down Time-Sensitive Clinical Decision Support

To further strengthen its diagnostic platform, the Client wanted to introduce a computer vision (CV) module that would automatically identify visual patterns in medical images, assisting in time-sensitive clinical decision-making. However, the new computer vision algorithm delivered by a third-party vendor showed insufficient accuracy (30%) during validation testing.

At the time, ScienceSoft was already providing L2-L3 support for the Client’s medical applications. While investigating recurring incidents, ScienceSoft’s engineers uncovered underlying usability, responsiveness, and code quality issues in the diagnostic platform. These issues slowed down clinical workflows and could affect the speed of medical decision-making.

To enhance the platform’s performance, the Client decided to revamp it, introduce new functionality, and refine the CV module before releasing it to physicians. Having already seen ScienceSoft’s proactive approach and engineering expertise during support cooperation, the Client chose ScienceSoft as a trusted technology partner for the platform evolution initiative.

ScienceSoft also stood out to the Client thanks to its profound experience in artificial intelligence and medical image analysis software, as well as its ISO 9001 and ISO 13485-certified quality management systems.

Diagnostic Platform Evolution and Image Analysis Module Refinement

ScienceSoft’s team began with a thorough review of the platform’s codebase. As a result, the Java engineers identified multiple areas for improvement in code quality and system responsiveness. Based on the assessment findings, they outlined a software evolution plan and implemented the necessary changes.

To improve the platform’s stability and performance, the development team conducted extensive code refactoring. Simultaneously, the engineers introduced several targeted upgrades to the platform’s functionality and usability, including:

• Introducing permanent access tokens and adding an authorization control system for third-party integrations to ensure their security.
• Integrating the platform with the internal systems of one of the Client’s partnering medical organizations to streamline data exchange.
• Replacing the unintuitive manual patient filtering interface with a more user-friendly design (with dropdown lists, checkboxes, and radio buttons).

On the computer vision front, ScienceSoft’s data scientists worked to boost the model’s diagnostic accuracy by focusing on three key aspects:

• Dataset enhancement: expanding the number of images annotated by experts, increasing dataset diversity, rebalancing class distribution, and optimizing tile size.
• Hyperparameter tuning: adjusting the batch size and the number of epochs to improve machine learning efficiency.
• Model fine-tuning: updating the model architecture to better accommodate real-world variability in clinical imaging data.

As the next step, the team built a custom image viewer to visualize CV output in an interactive format. It highlights areas of interest detected by the model, enabling medical professionals to quickly review the image and focus on regions that may require closer attention. Clinicians can also manually correct the annotations and submit them back to the model for further training.

After validating that the CV model’s accuracy has increased to an acceptable level (90%), ScienceSoft’s developers integrated the model with the platform’s back end and deployed it to production.

Improved Platform Usability and Tripled Medical Image Analysis Accuracy

As a result of engaging ScienceSoft, the Client saw a significant increase in the usability and responsiveness of its diagnostic platform used by physicians from partnering healthcare organizations to receive expert interpretation of medical images. The system now offers a smoother experience to end users working in fast-paced clinical environments, where quick access to information can speed up life-saving clinical decisions.

Most importantly, the revamp of the machine learning pipeline has tripled the accuracy of the medical image analysis model, increasing it from 30% to 90%. ScienceSoft also developed a custom medical image viewer for the platform. The viewer automatically highlights relevant image regions based on the model’s output, enabling medical specialists to provide instant feedback for further model training.

The functional upgrades implemented by ScienceSoft have also strengthened the platform’s security and streamlined data exchange between the platform and the partnering organizations’ internal systems.

Technologies and Tools

Java, JavaScript, React.js, TypeScript, Python, PyTorch, Amazon SageMaker, AWS S3, AWS Lambda