Release task
Inspect the task files, instructions, source material, and evaluation criteria.
Task requirements
Images load automatically. YAML loads when expanded; audio and other formats remain metadata-only.
Release task
Inspect the task files, instructions, source material, and evaluation criteria.
42 source files available in data/.
Task requirements
You are given eight coronary angiography keyframe cases. Each contains one primary keyframe plus four supplementary frames from the same acquisition and projection. Supplementary images show slight cardiac motion and contrast-filling changes and provide context for disambiguating overlap and catheter edges; they are not different projection angles.
For each case, identify visible coronary segments using SYNTAX conventions within the annotated keyframe, use same-acquisition supplementary frames to maintain vessel identity through motion and to resolve foreshortening/overlap, localize stenoses, and quantify severity using a mask-derived diameter rule, with registered evidence.
In the data directory:
case_01/ through case_08/ — one folder per case. Inside each:
case_01_keyframe.png — primary annotated keyframe (de-identified single-projection X-ray with contrast fill showing coronary tree: LM, LAD/diagonals/septals, LCx/OM, RCA branches depending on view, plus guiding catheter and aorta). This is the frame that held-back SYNTAX/stenosis annotations reference.case_01_frame_m2.png, case_01_frame_m1.png, case_01_frame_p1.png, case_01_frame_p2.png — four frames immediately before/after the keyframe from the same acquisition, same projection angle, same calibration. They show slight motion and contrast differences, not different projection angles (RAO/LAO). Use them to track vessels through motion and confirm true lesions vs overlap.cases.json — authoritative inventory: case ID, primary-keyframe basename, four supplementary-frame basenames, one projection label, pixel spacing, and dimensions. It contains no segment or stenosis answers.
segment_definitions.json — neutral SYNTAX segment dictionary listing every segment code that may appear, with anatomical names and definitions. It is a naming reference, not a per-case answer key.
Letter-suffixed source labels are merged into their anatomically identifiable
parent segment (for example, 16a and 16b become 16), so no hidden
annotation ordering is needed to name sibling branches.
The answer lives only in X-ray pixels of the primary keyframe, with same-acquisition frames as context. No multi-projection reconciliation — supplementary frames are same projection, same acquisition.
findings.json — per-case machine-readable inventory:{
"cases": [
{
"case_id": "case_demo",
"primary_keyframe": "case_demo_keyframe.png",
"supplementary_frames": [
"case_demo_frame_m2.png",
"case_demo_frame_m1.png",
"case_demo_frame_p1.png",
"case_demo_frame_p2.png"
],
"projection_label": "primary=-10 deg, secondary=25 deg",
"visible_segments": [
{
"segment_code": "06",
"segment_name": "LAD proximal",
"present": true,
"visible": true,
"keyframe": "case_demo_keyframe.png",
"confidence": 0.85,
"notes": "clear in keyframe, confirmed in supplementary m1/p1"
}
],
"stenoses": [
{
"id": "case_demo_s01",
"segment_code": "06",
"keyframe": "case_demo_keyframe.png",
"center_x": 212.3,
"center_y": 240.1,
"extent_px": 18.5,
"diameter_reduction_pct": 70.0,
"measurement_method": "mask-derived orthogonal local width versus segment 90th-percentile reference width",
"qca_geometry": {
"min_width_px": 3.6,
"reference_width_px": 12.0,
"min_width_line": [
[210.5, 240.1],
[214.1, 240.1]
],
"reference_center_x": 194.0,
"reference_center_y": 229.0,
"reference_width_line": [
[188.0, 229.0],
[200.0, 229.0]
]
},
"confidence": 0.8,
"uncertain": false,
"uncertainty_reason": null,
"confirmed_in_supplementary": true,
"temporal_track": [
{
"frame": "case_demo_keyframe.png",
"center_x": 212.3,
"center_y": 240.1,
"visible": true
},
{
"frame": "case_demo_frame_m2.png",
"center_x": 211.0,
"center_y": 241.0,
"visible": true
},
{
"frame": "case_demo_frame_m1.png",
"center_x": 210.8,
"center_y": 241.4,
"visible": true
},
{
"frame": "case_demo_frame_p1.png",
"center_x": 213.1,
"center_y": 239.8,
"visible": true
},
{
"frame": "case_demo_frame_p2.png",
"center_x": 214.0,
"center_y": 239.2,
"visible": true
}
]
}
],
"uncertain_regions": [
{
"segment_code": "11",
"keyframe": "case_demo_keyframe.png",
"reason": "catheter_overlap",
"bbox_xyxy": [180, 205, 238, 276],
"description": "guiding catheter overlaps proximal LCx in keyframe, persists in supplementary"
}
],
"catheter_notes": "JL4 in aortic root, overlaps proximal segments",
"overall_assessment": "Single 70% proximal LAD lesion in primary keyframe; other visible segments unobstructed, consistent across same-acquisition frames"
}
]
}
Required:
case_id must match cases.json. In findings.json, primary_keyframe, supplementary_frames, and every keyframe field must use file basenames only, exactly matching files inside that case folder. All supplementary frames are from the same acquisition and projection.visible_segments contains visible segments only; omit non-visible segments. Follow the bifurcation boundary rules in segment_definitions.json, reference the primary basename, and report confidence in [0,1].stenoses: id must be unique and use <case_id>_sNN; center_x/y use the original unresampled primary-keyframe pixel grid, origin (0,0) at top-left, x rightward and y downward. Define extent_px as the larger of the x-span and y-span of the connected narrowing region, measured between extreme pixel centers. For a quantifiable lesion, diameter_reduction_pct is 0-100, qca_geometry is required, and measurement_method must be exactly mask-derived orthogonal local width versus segment 90th-percentile reference width. If severity is unquantifiable, set uncertain:true, set both diameter_reduction_pct and qca_geometry to null, set measurement_method to exactly unquantifiable, and use one of catheter_overlap, overlap_foreshortening, poor_contrast, or border_uncertain. Unquantifiable lesions retain detection and uncertainty scoring but are excluded from severity scoring.min_width there. Define reference_width as twice the 90th percentile of all positive Euclidean distance-transform radii in the submitted segment mask; the reference line may be placed at any in-mask location having that width. Report 100 * (1 - min_width/reference_width), clipped to [0,100]. Use measurement_method: "mask-derived orthogonal local width versus segment 90th-percentile reference width".qca_geometry with positive min_width_px and reference_width_px, a two-endpoint min_width_line, a two-endpoint reference_width_line, and the reference-line center in original keyframe pixels. Each line length must agree with its declared width (within 1 pixel), its midpoint must be inside the submitted mask, and its direction must agree within 20 degrees of the local normal estimated from the closest mask boundary. The minimal-width line midpoint must agree with the stenosis center (within 3 pixels); the reference-line midpoint must agree with its declared center (within 3 pixels). The local mask width must agree with each declared width within 1.5 pixels, and the reported percentage must be derived from the two widths.temporal_track rows: the primary basename followed by the four supplementary basenames in cases.json order. For visible rows, give native-frame pixel coordinates. The primary row must be within 3 pixels of the submitted stenosis center, which itself must satisfy the lesion-region matching rule. For supplementary rows, the grader applies each held-back frame-to-frame motion delta to the submitted primary point and allows 12 pixels of correspondence tolerance, then also requires local-patch agreement with the submitted primary patch. This validates plausible same-lesion motion without forcing the held-back primary centroid. Filenames or booleans alone do not earn temporal credit.uncertain_regions per primary keyframe where catheter/overlap prevents assessment — check supplementary same-acquisition frames to confirm persistence. Every region requires bbox_xyxy: [x_min,y_min,x_max,y_max] in primary-keyframe pixels, with positive area and a description of at least 12 characters. The grader spatially compares the box with held-back catheter/border uncertainty geometry using three-pixel box expansion so tiny regions are not hypersensitive; prose alone earns no region credit.segment_masks/ — required registered vessel mapping:visible_segments, write segment_masks/<case_id>_<segment_code>.png.evidence/ (advisory and unscored):evidence/case_01_keyframe_overlay.png — overlay on primary keyframe per stenosis and per case showing segment labeling. Same dimensions as source keyframe. Draw centerlines, stenosis center+extent, label with segment codes and % per mask-derived rule.evidence/case_01_same_acquisition_sheet.png — contact sheet of primary keyframe plus supplementary same-acquisition frames (same projection) with same vessel/stenosis markers tracked across frames, proving temporal consistency and disambiguation of overlap/catheter vs true lesion.report.md (advisory and unscored) — temporal tracking, anatomy assignment, uncertainty, and severity notes.segment_definitions.json exact inventory. Maintain identity across supplementary same-acquisition frames of same case (same projection, slight motion). Non-visible segments not scored as false negatives.cases.json and segment_definitions.json.findings.json exists with all cases from cases.json, valid SYNTAX codes from segment_definitions.json, primary keyframe filenames per segment/stenosis, pixel coords, severity % per mask-derived orthogonal rule, confidence, uncertainty reasons, confirmation in supplementary same-acquisition frames.Images load automatically. YAML loads when expanded; audio and other formats remain metadata-only.
6 evaluation criteria
Visible coronary segments are assigned to the correct parent SYNTAX anatomy and spatially mapped with registered binary masks in each primary keyframe.
Each stenosis has a complete five-frame coordinate track whose local image patches support persistent identity.
Stenosis centers and extents match held-back annotations with false-positive findings penalized.
Diameter-reduction estimates include registered measurement-line geometry, follow the declared orthogonal-width rule, and agree with held-back mask-derived values.
Catheter overlap, poor contrast, and ungradable regions are handled as uncertainty rather than invented disease.
All declared case, anatomy, stenosis, uncertainty, and temporal fields are complete and use the authoritative frame basenames.