MIRO AI Research Interface

Applied AI / Frontend

Year: 2026
Built: DeepLabV3+ · API · UI
Context: Five-person research
Status: Non-diagnostic prototype

Designed orange MIRO cover reading AI medical imaging assistant above a standing person and a subtle cross-grid texture.

Overview

A research workspace for reviewing how a model maps the spine.

MIRO was a five-person master's research project. I built the DeepLabV3+ branch shown here, its FastAPI service, Next.js interface, and deployment. Teammates built the additional models and shared research pipeline.

A tour of MIRO, from its public experience and anatomy selection to the image-upload workspace.

Built with

Next.js 16
TypeScript
Tailwind CSS 4
GSAP
Three.js
FastAPI
PyTorch
DeepLabV3+

Research scope

Define the task before showing a result.

MIRO begins with the spine workflow, the only anatomy module available in this prototype. The task is clear before an image or model output appears.

MIRO anatomy selector with Spine selected on a standing person, an orange spine and pelvis overlay, and future anatomy modules shown as locked.

From input to review

One radiograph. Four working files.

Follow one approved demonstration radiograph from upload to result review. Each file shows one step of the same workflow.

Step 01

Step 01 / add image

Input

Add one approved radiograph.

The workspace holds the source image before MIRO runs the model. At this point, the page shows no result, label, or score.

MIRO / model-output review / 1 of 4

MIRO prediction workspace with a demonstration spine radiograph centered between thumbnail and zoom controls and a closed setup panel.

The protected workspace holds the source before any prediction is shown.

Non-diagnostic research prototype. Visible scores are model outputs—not diagnostic confidence or clinical accuracy.

Input

Add one approved radiograph.

The workspace holds the source image before MIRO runs the model. At this point, the page shows no result, label, or score.

Upload and results use separate states, making it clear when MIRO has produced an output.

This portfolio uses an approved demonstration image with no visible identifier.
No mask, polygon, label, or score appears before the model runs.

Configure

Choose what MIRO returns.

One choice controls what the model identifies. A second choice controls how the same result is drawn on screen.

Binary and multiclass change the model output. Raw and polygon change only its on-screen form.

Binary shows the spine as one region; multiclass separates vertebrae from T1 through L5.
Raw shows model pixels; polygon traces the same regions as shapes.

Binary: Shows the spine as one region against the background.
Multiclass: Separates T1 through L5 so each vertebra can be colored and labeled.
Raw: Shows the result as the pixels returned by the model.
Polygon: Traces those same regions as shapes that are easier to inspect.

Compare

See one result in two forms.

The radiograph and model output stay the same. Only the view changes—from raw pixels to outlined polygon shapes.

Switching views does not run the model again. It redraws the same output.

Raw shows the model's pixel-by-pixel mask.
Polygon traces the same regions as shapes that can display labels, layers, and boxes.

Raw mask: Direct access to the pixel-level model output.
Polygon overlay: Geometry organizes regions without changing the model mode.

Review

Review the full result or one vertebra.

The source image stays visible while controls reveal labels, layers, boxes, scores, and a closer vertebra view.

A researcher can move from the full model output to one vertebra without losing the original radiograph.

Labels, layers, boxes, and zoom change what is visible on screen.
Any visible score is a model output—not diagnostic confidence or clinical accuracy.

My implementation

The presented branch, end to end.

I carried the presented DeepLabV3+ work from model experiments to a deployed review interface.

01
Train and evaluate
DeepLabV3+ binary and T1–L5 multiclass experiments.
02
Serve the branch
Six FastAPI modes expose raw-mask and polygon outputs.
03
Protect the connection
A server-side Next.js proxy keeps backend details out of browser code.
04
Make outputs inspectable
Layers, labels, boxes, scores, zoom, and vertebra review.

Beyond the inference workspace.

Public pages, navigation, and additional review views complete the MIRO experience.

Drag to scroll

A research workspace for reviewing how a model maps the spine.

Define the task before showing a result.

One radiograph. Four working files.

Input

The presented branch, end to end.

Train and evaluate

Serve the branch

Protect the connection

Make outputs inspectable

Beyond the inference workspace.