Making Diagnostic Reasoning Addressable: The Work That Comes After DICOM

By Dmitry Golitsyn

AI can analyze information, but scaling expertise requires something deeper: infrastructure that makes human reasoning collaborative, traceable, and reusable.

Most organizations view AI adoption as a technology challenge. In reality, the larger obstacle is often architectural. AI can only create meaningful value when human expertise, decisions, and context can be captured, shared, and revisited within structured workflows. Drawing on examples from radiology, software development, and cloud infrastructure, Dmitry Golitsyn explores why the next stage of digital transformation is not simply about smarter algorithms, but about making expertise itself addressable, collaborative, and scalable.

The first few minutes of any radiology second-opinion exchange are nearly always the same. The radiologist asking for input has to manually reconstruct the entire clinical context for the colleague joining the case: which study, which series, which slice, which window setting, which prior, and the exact region they want a second look at.

The imaging data itself is already shared across the network effortlessly, but the diagnostic state around it remains completely trapped.

This gap seems small when looking at any single case. But across an entire department or an enterprise-scale health system, this friction compounds rapidly.

Because our systems cannot natively track this fragmented data, human collaboration is forced into informal side-channels, and AI assistance is left stranded – sitting awkwardly beside the workflow rather than inside it.

Fixing this is not a clinical problem; it is a fundamental architecture problem. The next infrastructure layer in radiology must bridge this gap by making diagnostic context directly addressable. This is the exact paradigm shift driving DICO, an enterprise platform engineered to turn fragmented clinical workflows into a connected, navigable network of immutable image addresses.

In this article, I look under the hood of DICO’s architecture to explore how native addressability unifies peer reviews, deeply integrates AI support, and secures the future of medical imaging informatics.

The Retrieval Baseline: Why Storage Is Not Collaboration

The medical imaging industry did the heavy infrastructure work decades ago. DICOM standardized how imaging objects are stored, identified, transferred, and retrieved across disparate hospital systems. Every study, series, and instance has a unique identifier, and DICOMweb eventually extended that framework to web-native access. But simple data retrieval is not the same as active collaboration.

The friction exists because our standard infrastructure operates on the scale of massive, multi-gigabyte files, while human collaboration happens at an entirely different resolution.

For a clinical expert, the actual unit of diagnostic reasoning is much smaller and more granular than an entire image series. It is a subtle, suspicious lesion on a specific slice. It is a precise measurement compared directly against a historical prior scan. It is the exact reconstruction plane where a confusing artifact finally disappears, a specific region that requires subspecialty review, or a single sentence in the final report that depends on an isolated image observation.

Today, almost all of that intricate reasoning lives entirely in the radiologist’s working memory. It is highly visible during the fluid moment of interpretation, but it is compressed, flattened, and largely lost the moment the final report is signed.

A colleague joining the case even an hour later cannot easily point to “the exact thing I was unsure about” because the system provides no unique digital address for it.

The legacy address space, study → series → instance, stops short of reaching inside the actual diagnostic question. A functional, modern address must extend further and include frame, spatial region, viewport state, clinical question, contribution status, and provenance.

Without this granular addressability, collaboration is forced out of the imaging system and into fragmented side channels like screenshots, phone calls, and hallway consults. This happens because legacy medical software still treats a digital image the same way we used to treat physical film: as a static, closed object that can only be viewed by one person at a time.

The Collaborative Blueprint: Moving Beyond Static Files

To see how we can break out of this workflow trap, the best model for this structural shift is not found in legacy medical software but in modern web-collaborative tools like Figma.

Figma fundamentally transformed the design industry not just by building slicker vectors but by changing the basic nature of the file itself. A design file stopped being a static asset that one person edited in isolation and others reviewed later. It became a live, shared workspace where multiple people could see, comment, edit, and coordinate around the exact same digital object in real time. While Figma synchronizes this document state seamlessly using web sockets and multiplayer servers, the deeper conceptual breakthrough was treating the file as a highly structured environment built from addressable, interactive objects.

A complex CT scan is obviously not a design canvas. Medical imaging involves massive volumetric data, contrast phases, and strict layers of clinical responsibility, regulation, and legal accountability that Figma never has to navigate. This transition is genuinely harder, not just slower.

But the underlying product lesson is completely portable: collaboration only becomes natural when the object of work becomes shared and addressable. In radiology, that object should be the imaging study – not viewed as a static file, but utilized as a dynamic diagnostic workspace.

The Five Pillars of an Addressable Imaging Architecture

A functional collaboration layer cannot just be a cosmetic user-interface feature added to a standard viewer. It should operate as an infrastructure layer that binds together image data, human expertise, AI outputs, and the final report. At the bare minimum, it requires five core capabilities:

1. Persistent Diagnostic Anchors (Asynchronous State). Comments, measurements, and AI suggestions cannot float loosely outside the study. They must be anchored to an immutable visual context – including the exact series, slice, coordinate region, window settings, and zoom level. When a clinician leaves a note for a colleague, clicking that note must reconstruct that exact visual state layout instantly, ensuring the asynchronous reviewer sees precisely what the author saw.
2. Multi-Player Viewport Sync (Synchronous Collaboration). True real-time collaboration fails when two specialists believe they are discussing the same pathology but are actually looking at slightly different slices or window settings. Beyond saving static anchors, a web-native collaborative environment allows a radiologist to actively broadcast their focus – syncing live viewports, cursor movements, and scrolling sequences across distant monitors so that shared attention becomes completely instantaneous.
3. Semantic Regions (Data Enrichment). A true diagnostic object possesses a type, a status, an author, a confidence level, and an explicit relationship to priors. This is where AI becomes immensely useful without overstepping into the role of a physician – by automatically labeling anatomy, suggesting candidate findings, and transforming raw pixels into structured, searchable data regions.
4. Contribution Provenance (The Assembly Process). When a complex diagnosis is assembled from multiple subspecialty inputs, the system must preserve the strict version history of every data point. The architecture needs to track who created a measurement, who edited it, and whether it originated from an AI model. Scoped contributions tied to identifiable image regions give radiology the same precise, multi-author audit trail that version control gives to software engineering.
5. Report Traceability (The Final Deliverable). The final text report should never be an isolated text file disconnected from the visual evidence that produced it. Report statements must link directly back to the underlying semantic regions and accepted peer reviews. While the final text remains concise for referring physicians, the interactive diagnostic graph is preserved underneath it, making every conclusion explicitly traceable back to the image.

Redefining the Interface: From a Viewer to a Workspace

To actually realize these five capabilities, we have to rethink the very nature of radiology software.

The radiologist’s traditional tool is no longer adequate when this collaborative mindset is adopted. The word “viewer” is simply becoming too small for modern medicine. A viewer merely displays static images; a diagnostic workspace actively coordinates the human work occurring around them.

Most modern radiology departments are already trying to force this collaborative model into existence through informal workarounds. Curbside consults, double-reads, and rapid subspecialty referrals are examples of the clinical workflow attempting to be collaborative without the necessary underlying infrastructure to support it.

The high cost of that technical gap lands squarely on the working radiologist, who is forced to carry complex diagnostic states entirely in their short-term memory because the software cannot.

To relieve this cognitive burden, the underlying architecture should change from static file storage to a dynamic database. In a modern, addressable architecture, the imaging study becomes the active center of the workflow. Every single finding, measurement, clinical question, AI suggestion, and specialist review attaches to the exact same addressable space.

When the software itself can track and store this granular context, it fundamentally redefines how we interact with technology. This is what a genuinely “AI-native” environment looks like in radiology. The human radiologist remains the sole decision-maker, while AI works alongside them on the exact same diagnostic objects, operating with clear roles, clear provenance, and absolute accountability.

The Next Infrastructure Layer: Putting Addressability into Practice

Radiology has strong infrastructure for storing, transmitting, and displaying images. Building that data-delivery foundation took decades, and it was a mandatory first step.

The next layer makes diagnosis itself navigable. From private interpretation to structured contribution. From screenshots to precise image addresses. From informal second opinions to traceable review. From AI outputs beside the workflow to AI support inside it.

The payoff is immediate. We stop losing the clinical logic that currently vanishes the moment a case is closed. At the same time, we end the constant time drain of manually hunting down coordinates and context for every second opinion.

Diagnosis is assembled. The infrastructure around DICOM should be built for that.

About the Author

Dmitry Golitsyn is Co-Founder at DICO by Expert Radiotech, a medical technology company building a new operating model for diagnostic radiology. His work focuses on healthcare infrastructure, radiology workflows, AI-enabled collaboration, and the future of expertise in complex medical systems. He writes about how workflow architecture and distributed intelligence can improve scalability, quality, and access to diagnostic care.