feat(ceus): 两半坐标换算改为每段视频配准，加目视核对脚本

- 第 1 步分屏后在两半扇形区内用边缘图归一化互相关估 (dx, dy)，再用逐帧多块模板匹配的中位数独立核对；两者各分量之差超过门槛（15 像素）这一例第 1 步失败。删掉「两个视图左上角相减」的实现与测试。 - 门槛依据：34 例中走完第 1 步的 29 例，差值中位 2、95 分位 6.2、最大 13；参数与门槛在同一批视频上调出。 - scripts/ceus_align_review.py：只做第 1 步，出扇形区棋盘图与边缘叠加图，混入错开 15–30 像素的阴性对照、随机编号；计分时任一对照未被判「没对齐」即作废。 - ceus_quant 读 review.csv：不通过的病例第 1 步失败；表里没有这段视频或平移对不上即停。 - 配准与紫圈颜色参数进分割缓存键。 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-16 06:47:27 -04:00
parent c582fe426f
commit 00ebbffb28
20 changed files with 1500 additions and 277 deletions
@@ -0,0 +1,196 @@
+"""`scripts/ceus_align_review.py`：配准目视核对的出图和计分（任务书 02「配准的目视核对」）。
+
+出图：假交付包两例合成分屏视频（`fake_ceus_video`，两半扇形对齐，配准 (0, 0)），对照取 1 例。
+计分：key.csv 和两份判读手写，预期按任务书的规则手算。表头按任务书和协调者给的列名独立写。
+"""
+
+import csv
+import hashlib
+import importlib.util
+import math
+import re
+from pathlib import Path
+from types import ModuleType
+from typing import Any
+
+import cv2
+import pytest
+import yaml
+
+from conftest import CEUS_YAML
+from fake_ceus_video import write_video
+from fake_delivery import make_delivery
+
+_SCRIPT = Path(__file__).resolve().parents[1] / "scripts" / "ceus_align_review.py"
+_KEY_HEADER = ["id", "video_id", "is_control", "dx", "dy", "video_sha256"]
+_VERDICT_HEADER = ["编号", "判定", "依据"]
+_REVIEW_HEADER = ["video_sha256", "dx", "dy", "passed", "verdict_a", "verdict_b"]
+
+
+def _load_script() -> ModuleType:
+    spec = importlib.util.spec_from_file_location("ceus_align_review", _SCRIPT)
+    assert spec is not None and spec.loader is not None
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+def _read(path: Path) -> list[list[str]]:
+    with path.open(encoding="utf-8", newline="") as handle:
+        return list(csv.reader(handle))
+
+
+def _write(path: Path, rows: list[list[str]]) -> None:
+    with path.open("w", encoding="utf-8", newline="") as handle:
+        csv.writer(handle).writerows(rows)
+
+
+def test_render_writes_shuffled_images_controls_and_key(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch, capsys: pytest.CaptureFixture[str]
+) -> None:
+    delivery_dir = make_delivery(2).write(tmp_path / "delivery")
+    videos = {
+        "U001": delivery_dir / "CEUS_agent" / "001" / "U001_roi2.mp4",
+        "U002": delivery_dir / "CEUS_agent" / "002" / "U002_roi2.mp4",
+    }
+    write_video(videos["U001"], green_box=True)
+    write_video(videos["U002"], green_box=False)  # 第 2 步才会失败，第 1 步照常出图
+    document: dict[str, Any] = yaml.safe_load(CEUS_YAML.read_text(encoding="utf-8"))
+    document["paths"].update(
+        delivery_dir=str(delivery_dir),
+        align_runs_root=str(tmp_path / "align"),
+        align_review=str(tmp_path / "unused_review.csv"),  # 出图用不到，文件不存在也行
+    )
+    document["align_check"]["control_count"] = 1
+    config = tmp_path / "ceus.yaml"
+    config.write_text(yaml.safe_dump(document, allow_unicode=True), encoding="utf-8")
+    module = _load_script()
+
+    def fake_run_id(now: object, git: object) -> str:
+        return "run-a"
+
+    monkeypatch.setattr(module, "make_build_id", fake_run_id)
+    module.main(["render", "--config", str(config)])
+
+    run_dir = tmp_path / "align" / "run-a"
+    assert not (tmp_path / "align" / "run-a.partial").exists()
+    key = _read(run_dir / "key.csv")
+    assert key[0] == _KEY_HEADER
+    rows = key[1:]
+    # 2 例真实 + 1 例对照，编号是 001–003 的一个排列，每个编号两张图。
+    assert sorted(row[0] for row in rows) == ["001", "002", "003"]
+    names = sorted(p.name for p in (run_dir / "images").iterdir())
+    assert names == sorted(f"{row[0]}_{kind}.png" for row in rows for kind in ("checker", "edges"))
+    assert all(re.fullmatch(r"\d{3}_(checker|edges)\.png", name) for name in names)
+    real = [row for row in rows if row[2] == "false"]
+    controls = [row for row in rows if row[2] == "true"]
+    assert sorted(row[1] for row in real) == ["U001", "U002"]
+    assert [(row[3], row[4]) for row in real] == [("0", "0"), ("0", "0")]
+    (control,) = controls
+    # 对照的平移 = 真实配准 (0, 0) + 长度 15–30 像素的随机错开（取整后容差 1 像素）。
+    assert 14 <= math.hypot(int(control[3]), int(control[4])) <= 31
+    for row in rows:
+        assert row[5] == hashlib.sha256(videos[row[1]].read_bytes()).hexdigest()
+    checker = cv2.imread(str(run_dir / "images" / f"{control[0]}_checker.png"))
+    edges = cv2.imread(str(run_dir / "images" / f"{control[0]}_edges.png"))
+    assert checker is not None and edges is not None and checker.shape == edges.shape
+    assert _read(run_dir / "failed.csv") == [["video_id", "step", "reason"]]
+    assert (run_dir / "snapshot" / "ceus.yaml").read_bytes() == config.read_bytes()
+    out = capsys.readouterr().out
+    assert "U001" not in out and "U002" not in out
+
+
+# 手写的 key.csv：两例真实（001、003）、一例对照（002）。
+_KEY = [
+    _KEY_HEADER,
+    ["001", "V1", "false", "-24", "0", "a" * 64],
+    ["002", "V1", "true", "-4", "17", "a" * 64],
+    ["003", "V2", "false", "183", "-2", "b" * 64],
+]
+
+
+def _score(module: ModuleType, run_dir: Path, a: list[list[str]], b: list[list[str]]) -> None:
+    _write(run_dir / "key.csv", _KEY)
+    _write(run_dir / "verdicts_a.csv", [_VERDICT_HEADER, *a])
+    _write(run_dir / "verdicts_b.csv", [_VERDICT_HEADER, *b])
+    module.main(["score", "--run-dir", str(run_dir)])
+
+
+def test_score_writes_review_when_controls_caught(
+    tmp_path: Path, capsys: pytest.CaptureFixture[str]
+) -> None:
+    """对照两人都判「没对齐」；001 两人都判对齐 → 通过，003 有一人看不清 → 不通过。"""
+    module = _load_script()
+    a = [["001", "对齐", "血管连续"], ["002", "没对齐", "边缘错开"], ["003", "对齐", "扇形重合"]]
+    b = [["001", "对齐", "血管连续"], ["002", "没对齐", "边缘错开"], ["003", "看不清", "太暗"]]
+    _score(module, tmp_path, a, b)
+
+    assert _read(tmp_path / "review.csv") == [
+        _REVIEW_HEADER,
+        ["a" * 64, "-24", "0", "true", "对齐", "对齐"],
+        ["b" * 64, "183", "-2", "false", "对齐", "看不清"],
+    ]
+    out = capsys.readouterr().out
+    assert "通过 1，不通过 1" in out
+    assert "「看不清」1 例" in out and "不一致 1 例" in out
+    # ceus_quant 读得回这份表。
+    review = _load_quant().read_review(tmp_path / "review.csv")
+    assert review["a" * 64].offset == (-24, 0) and review["a" * 64].passed is True
+    assert review["b" * 64].passed is False
+
+
+def _load_quant() -> ModuleType:
+    spec = importlib.util.spec_from_file_location("ceus_quant", _SCRIPT.parent / "ceus_quant.py")
+    assert spec is not None and spec.loader is not None
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+@pytest.mark.parametrize("control_verdict", ["对齐", "看不清"])
+def test_score_is_void_when_a_control_is_missed(
+    tmp_path: Path, capsys: pytest.CaptureFixture[str], control_verdict: str
+) -> None:
+    module = _load_script()
+    a = [["001", "对齐", "x"], ["002", "没对齐", "x"], ["003", "对齐", "x"]]
+    b = [["001", "对齐", "x"], ["002", control_verdict, "x"], ["003", "对齐", "x"]]
+    with pytest.raises(SystemExit) as raised:
+        _score(module, tmp_path, a, b)
+    assert raised.value.code == 1
+    assert not (tmp_path / "review.csv").exists()
+    assert "作废" in capsys.readouterr().out
+
+
+@pytest.mark.parametrize(
+    ("rows", "message"),
+    [
+        ([["001", "对齐", "x"], ["002", "没对齐", "x"]], "缺"),  # 缺 003
+        (
+            [
+                ["001", "对齐", "x"],
+                ["002", "没对齐", "x"],
+                ["003", "对齐", "x"],
+                ["004", "对齐", "x"],
+            ],
+            "多",
+        ),
+        ([["001", "对齐", "x"], ["002", "没对齐", "x"], ["003", "差不多", "x"]], "判定只能是"),
+        (
+            [
+                ["001", "对齐", "x"],
+                ["001", "对齐", "x"],
+                ["002", "没对齐", "x"],
+                ["003", "对齐", "x"],
+            ],
+            "重复",
+        ),
+    ],
+)
+def test_score_rejects_malformed_verdicts(
+    tmp_path: Path, rows: list[list[str]], message: str
+) -> None:
+    module = _load_script()
+    good = [["001", "对齐", "x"], ["002", "没对齐", "x"], ["003", "对齐", "x"]]
+    with pytest.raises(ValueError, match=message):
+        _score(module, tmp_path, rows, good)
+    assert not (tmp_path / "review.csv").exists()