In-browser inference of the Mirai
model via ONNX Runtime Web. Four mammogram DICOMs go in, five calibrated year-hazards plus a 612-dim explainability embedding come out — no data leaves the browser.
1. Input
Drop 4 DICOMs here (L-CC, R-CC, L-MLO, R-MLO) or
or
2. Preview
View and laterality auto-detected from each DICOM's ViewPosition (0018,5101) and ImageLaterality (0020,0062). Pixels are 1st–99th percentile stretched for display only — the model receives the full-range tensor.
3. Risk factors (optional)
Listed in the canonical order of the 34 risk-factor keys (matches the 100-dim RF block of the embedding). Missing keys fall back to the model-predicted risk-factor path.
shortcut if specifics unknown
auto-derived from age & menopause age
Note: setting either BC or OC for a relative marks the other as "no" if left "unknown" — upstream Mirai ties known-status per relative.
4. Run
EP: —threads: —
Waiting for DICOMs…
Predictions
year
rounded (4dp)
calibrated (fp64)
Slot order
slot
view
side
flipped
XAI embedding (612-dim, post-ReLU)
Laid out as 34 rows × 18 cols (= 612). Darker = larger activation. First 512 dims are image features (img_emb_1…512); last 100 are the risk-factor block (density_1…hrt_years_ago_stopped_gt_7). Hover any cell for its label and activation.