Why SpatialX: Reach the Pathologist

Before any architecture, a few plain definitions, because the rest of the course leans on them. A pathologist is the doctor who diagnoses disease — including cancer — by examining tissue under a microscope. A whole-slide image (WSI) is that same glass microscope slide scanned into one enormous digital picture. Gigapixel means the picture has billions of pixels — a single scan can be roughly 100,000 by 100,000 pixels. Diagnosis is naming what a disease is; prognosis is predicting how it will progress.

Keep one mental picture for the whole course: a doctor looking at a giant digital slide, and software that has to help without getting in the way.

The diagram is the whole product in one line: a slide becomes a WSI, the WSI becomes tiles the AI and the viewer can handle, the AI proposes labels, and the pathologist reviews and corrects them on one screen. Every later day is a zoom-in on one of these boxes.

The starting point is a product problem, not an architecture problem. Cancer diagnosis today runs on gigapixel images, and even when a genuinely good AI model exists — in a research lab, in a paper, on a benchmark — it rarely makes it into the hands of the pathologist at the moment they are actually diagnosing a patient. The model and the doctor live on opposite sides of a gap.

The box-art shows the gap literally: a working model on the left, the doctor who needs it on the right, and nothing that reliably connects them. Models are trained and published all the time; almost none are wrapped into something integrated, secure, and usable at the point of care. Closing that gap is the product.

Orca is a complete, end-to-end AI platform for cancer diagnosis and prognosis. "End-to-end" means it covers the whole path — getting a slide in, tiling it, running AI on it, letting a pathologist annotate and correct, and storing the result — not just the model in the middle. To operationalise ML means to turn a research model into a dependable product feature: something with a user interface, security, and error handling around it, so a person who is not a machine-learning expert can rely on it.

The key user constraint: pathologists are experts in tissue, not in machine learning. So the platform has to hide the ML machinery entirely and present something that feels like a familiar medical tool — integrated into their workflow and secure by default.

The diagram is the reframe: the value is not the model alone, it is everything Orca puts around the model so a non-ML expert can trust it. The accuracy is table stakes; the delivery is the product.

This gap is worth closing because of its scale and its direction. CAGR below means compound annual growth rate — the steady yearly percentage a market grows. The numbers come from the deck's own market slide.

The two forces point in opposite directions, which is exactly what makes the problem urgent.

The diagram makes the tension concrete: demand climbs while supply falls, so the number of slides each pathologist must read keeps rising. Software that lets one pathologist safely do more is not a luxury in that world — it is the only way the arithmetic works.

One honest note, because it is the real reason the work exists rather than a market slide. I lost my mom to cancer in 2015. That is why I worked on SpatialX. The market numbers explain why the problem matters at scale; this is why it mattered enough to build.

One idea runs under every decision in this course, so it is worth stating on day one. The platform was built so that the customer, and the small team building it, could bend it without the senior engineer in the loop for every change. Put differently: the goal of every architectural choice you will see was to take the senior engineer out of the critical path.

The diagram is the strategy in miniature: most new work should be a configuration change against an existing pattern, not new architecture. When something genuinely new is needed, a senior builds it once and it, too, becomes a template. Day 2 turns this idea into the concrete plan that let a two-to-three-person team ship a gigapixel cancer-AI platform.

• Orca is an end-to-end AI platform for cancer diagnosis and prognosis: it takes a slide all the way from upload, through tiling and AI inference, to a pathologist reviewing and correcting annotations on one screen.
• The product problem comes first: cancer diagnosis runs on gigapixel whole-slide images, and even good AI models rarely reach the pathologist at the point of diagnosis. Closing that gap is the whole point.
• The value is not the model alone — it is operationalising the model: wrapping it in a UI, security, and error handling so a non-ML expert can depend on it.
• The stakes are large and worsening: ~20M new cancer cases and ~9.7M deaths a year (WHO/IARC GLOBOCAN 2022), demand up ~40% by 2035 against a shrinking pathologist workforce, in a digital-pathology market growing $1.46B → $2.75B by 2030 (13.5% CAGR).
• The throughline for every later day: build the system so the customer and the team can bend it without the senior engineer — take the senior out of the critical path.

• [ ] I can define pathologist, whole-slide image (WSI), gigapixel, diagnosis, and prognosis in one sentence each.
• [ ] I can state Orca's product problem: the best AI never reaches the pathologist.
• [ ] I can explain what "operationalise ML for non-ML experts" means and why the model alone is not the product.
• [ ] I can recite the core market figures and explain why rising demand plus a shrinking workforce makes the problem urgent.
• [ ] I can state the throughline — take the senior engineer out of the critical path — and why day 2 builds on it.