Radiology Browser Federated Demo

Research/demo boundary

This prototype is not for diagnosis, not a medical device, and not for patient-care decisions.

Phase008 runs local lung segmentation, derives reproducible dual-view preprocessing contract, and can send tiny adapter weights/metrics only to the local coordinator. It does not recognize disease, upload pixels, or train the full U-Net.

Local model, image, and optional local mask

X-ray image file Decoded locally with browser canvas, resized to 256×256 grayscale 0..1. Optional local binary mask Used only for local adapter calibration. Raw mask is not exported. Optional local ONNX model Default path is models/phase136w_student_tiny_unet_nobn.onnx. Use picker if the static file is absent.

Waiting for model and image.

Research input-quality gate

Deterministic engineering checks only. Projection is selected by the user and is not inferred. Passing does not establish clinical image quality.

Declared projection

gate result: not assessed
source dimensions: not available
grayscale mean: not available
grayscale std: not available
dark clip fraction: not available
bright clip fraction: not available
mask border contact: not available

Warnings: Load an image to run local deterministic checks.

Input preview

Preview only. The model input is internally resized to 256×256.

Mask / adapter overlay

Overlay from local ONNX output, or adapter-calibrated output after local adapter training.

Lung-focused classifier preview

Derived locally from the cleaned segmentation mask. This is preprocessing only; no disease classifier is loaded or executed.

Classifier preprocessing evidence

Both pathways are preserved. Images use bilinear resize with aspect-preserving letterbox padding. Masks use nearest-neighbor geometry only. Pixels remain local.

Lung crop margin 0.04 Normalization

preprocessing status: waiting for segmentation
contract version: radiology-dual-view-v2
lung bbox: not available
padded bbox: not available
crop area fraction: not available
full-view letterbox: not available
crop letterbox: not available
classifier tensor size: 256×256×1 NHWC
image interpolation: bilinear
mask interpolation: nearest-neighbor
disease inference: false

Postprocessing

Threshold 0.75

largest-two-components cleanup show overlay show adapter output when available

mask_area_fraction: not available yet
runtime: onnxruntime-web
local_only: true
raw_upload: false

Tiny adapter training

Trains only q = sigmoid(bias + scale_p*p + weight_x*x + weight_y*y + weight_r*r) over the frozen ONNX probability map. This is not full U-Net backprop.

Epochs

Learning rate

mask loaded: false
truth_area_fraction: not available
adapter dice before/after: not trained
adapter loss before/after: not trained

Open coordinator

worker_id: not joined
coordinator transport: BroadcastChannel
last send status: not sent

Install-less browser radiology AI demo