Starting Point: Complete Domain Unfamiliarity
The project began as an R&D initiative to build a proof-of-concept for a medical imaging AI orchestration platform. The technical brief mentioned Mercure and MONAI as core technologies. On paper, it appeared to be a standard development project.
The reality was far more complex.
Coming into this project with no previous exposure to medical imaging, the domain presented immediate challenges. The first hurdle wasn't architectural or technical—it was understanding the fundamental terminologies. What exactly is a "lesion"? MS lesions, as it turns out, are areas of damage or scarring in the central nervous system associated with Multiple Sclerosis. These small spots visible on brain scans carry critical diagnostic significance that wasn't immediately obvious to someone outside the medical field.
Understanding DICOM
The first major technical challenge was understanding DICOM—the Digital Imaging and Communications in Medicine standard. Initial research led to several key resources, including the "Introduction to DICOM" blog series, PyDICOM documentation, and the official DICOM standard documentation from NEMA. These resources revealed that medical images aren't simple picture files; they're complex data structures containing patient information, acquisition parameters, and metadata crucial for diagnosis.
Working with DICOM files programmatically using PyDICOM showed the difference between medical imaging and conventional image formats. Unlike JPEG or PNG files, DICOM files carry everything from pixel data to patient demographics, study descriptions, and technical parameters. Each DICOM file contains hundreds of tags, each serving a specific purpose in the medical workflow.
Understanding the structure took considerable time. ContentSequence tags, Structured Reports, Study Instance UIDs—terms that initially appeared as technical jargon but gradually revealed their significance in the medical imaging ecosystem. The complexity of DICOM becomes apparent when you realize it's not just an image format, but a comprehensive standard for medical image communication and storage.
Workflow Orchestration with Mercure
Mercure served as the introduction to medical imaging workflows and orchestration. The initial setup involved configuring Amazon EC2 instances, managing Docker containers, and understanding how different components communicate in a hospital environment.
Mercure's role extends beyond simple image processing—it manages complex workflows. When a DICOM study arrives, Mercure can automatically trigger AI analysis, route results to different systems, and generate reports while maintaining the strict standards required in healthcare.
The infrastructure setup included working with dcmsend commands and managing the 60-second reception timeout, which makes sense when you understand that medical studies often consist of hundreds of images that need to be received completely before processing can begin.
Initial research into DICOM server solutions led to evaluating Orthanc, a lightweight DICOM server. However, for our specific requirements, Orthanc seemed heavy-handed given our minimal use case and the need for tight integration with our existing database and backend technology stack.
Instead, we built a custom DICOM server using pynetdicom. This approach provided several advantages: direct integration with our database architecture, alignment with our backend technology stack, and precise control over the DICOM operations we needed to support.
The pynetdicom library offered the flexibility to implement exactly the DICOM services required without the overhead of a full-featured server like Orthanc. This custom implementation allowed seamless integration with our workflow orchestration and maintained the critical data integrity standards essential in medical imaging, where every pixel matters for radiological diagnosis.
Medical Image Viewing and Custom Development
Working with OHIF (Open Health Imaging Foundation) and Cornerstone provided the foundation for understanding medical image visualization. These platforms demonstrated the sophistication required for medical image interpretation beyond simple image display.
Cornerstone's rendering capabilities highlighted the complexity of medical image display. Windowing and leveling, which adjust brightness and contrast to highlight specific tissue types, became essential concepts.
Our specific requirements for lesion tracking led to the development of a custom solution from scratch. After evaluating OHIF's customization capabilities, it became clear that achieving the precise lesion tracking functionality needed would require building a custom lesion tracker rather than adapting existing OHIF components.
The custom lesion tracker development involved working directly with Cornerstone's rendering engine while building specialized tools for MS lesion identification, measurement, and tracking across multiple time points. This approach provided the granular control necessary for the specific clinical workflow requirements.
Medical AI with Open Source
Open source frameworks have transformed medical imaging AI by providing specialized tools for healthcare's unique challenges. MONAI's comprehensive framework, TotalSegmentator for automated organ segmentation, and the MONAI Model Zoo's pre-trained models eliminate the need to build from scratch while ensuring clinical-grade quality.
Medical AI demands more than accuracy—it requires reliability and interpretability. A false positive in tumor detection can lead to patient anxiety and unnecessary procedures. Open source medical AI tools address these stakes through robust validation, standardized pipelines, and transparent architectures that clinicians can understand and trust.
This ecosystem accelerates innovation while maintaining the rigorous standards essential for healthcare applications.
Domain Knowledge Acquisition
The technical learning curve was paralleled by necessary medical domain knowledge acquisition. Understanding medical terminology became essential for effective platform development. Axial, sagittal, and coronal views became standard concepts, along with understanding why certain imaging protocols exist and how different modalities—CT, MRI, PET—serve different diagnostic purposes.
The learning process involved understanding why DICOM Structured Reports exist, how to generate PDF reports that integrate with hospital workflows, and ensuring that AI-generated insights are presented in formats that clinicians can trust and act upon.
Systems Integration Perspective
The most significant realization was that medical imaging isn't about individual technologies—it's about integrated ecosystems. Mercure orchestrating workflows, MONAI providing AI capabilities, OHIF and custom viewers enabling visualization, and DICOM ensuring interoperability. Each component serves a critical role in delivering patient care.
The initial technical challenges—DICOM tag complexity, medical terminology, strict data handling requirements—all served purposes within a larger system designed to support healthcare delivery.
Conclusion
This project demonstrated that building for healthcare requires accepting higher standards of responsibility. Every technical decision potentially impacts patient outcomes, requiring a different approach to software development than typical commercial applications.
The complexity that initially presented significant challenges—from understanding MS lesions to configuring AI orchestration pipelines—provided appreciation for the sophisticated systems supporting modern medicine.
Medical imaging technology operates at the intersection of advanced AI, complex data management, and life-critical decision making. It's a domain where technical excellence extends beyond performance metrics to enabling better patient care.
For technical professionals considering healthcare technology, the domain knowledge requirements are substantial, but the potential impact is significant. Every successful DICOM transfer, accurate AI prediction, and properly rendered medical image contributes to systems that improve patient outcomes.
The complexity exists for important reasons, and mastering it enables meaningful contributions to healthcare technology advancement.
This write-up is a future reference to myself and to anyone who’s getting started with the development of a medical imaging platform.
