hkust-aaron's picture
Physician review recorded; updated labels metadata and showcase videos
b91b0f5 verified
Raw History Blame Contribute Delete
18.6 kB
"""Task hierarchy, symbolic state and decision points for the three action-labelled clips.
Usage: python 17_task_graph.py <pack_dir>
Adds the "why" layer on top of layer5_actions/actions.json:
goal -> subtasks -> action segments, where each subtask has
preconditions / effects symbolic predicates over instruments and tissue,
e.g. held(gauze, #101), under_tension(material)
success_check what is visible when the subtask has worked
decision points: observation -> assessment -> decision -> expected result,
each anchored to a frame, and the outcome actually seen afterwards.
Written by an AI model from the video. Observations and outcomes describe only
what is visible; assessments and decisions are a plausible rationale for what
the operator did, not a record of their intent. Reviewed and approved by the contributing physician (2026-11-17).
Writes layer5_actions/task_graph.json.
"""
import json
import re
import sys
from pathlib import Path
pack = Path(sys.argv[1])
acts = json.load(open(pack / "layer5_actions" / "actions.json"))
PREDICATES = {
"held(X, I)": "object X is between the closed jaws of instrument I",
"on(X, Y)": "object X lies on / presses against Y",
"visible(X)": "X can be seen in the view",
"jaws_open(I)": "instrument I's jaws are open",
"between_jaws(X, I)": "X lies between the open jaws of I",
"under_tension(X)": "X is stretched by traction",
"separated(X, Y, p)": "X has come off Y; p = none / partial / complete",
"energy_applied(X)": "smoke, darkening or char appears at X",
"out_of_view(I)": "instrument I is not in the image",
"red_spot_visible(X)": "a red spot / ooze is visible on X",
}
T = {
5: {
"goal": "Control oozing above the lobules with gauze pressure",
"initial_state": ["held(gauze, #101)", "visible(oozing_surface)", "red_spot_visible(oozing_surface)", "on(#102, tissue_right)"],
"subtasks": [
{"id": "5.1", "name": "Bring the gauze to the oozing area", "segments": [[101, 1250], [101, 1263]],
"pre": ["held(gauze, #101)"], "effects": ["gauze over oozing_surface"],
"success_check": "gauze edge reaches the upper oozing area"},
{"id": "5.2", "name": "Apply pressure", "segments": [[101, 1275]],
"pre": ["gauze over oozing_surface"], "effects": ["on(gauze, oozing_surface)"],
"success_check": "gauze stays on the area for about 3 s"},
{"id": "5.3", "name": "Check underneath", "segments": [[101, 1346]],
"pre": ["on(gauze, oozing_surface)"], "effects": ["visible(oozing_surface)"],
"success_check": "the area is uncovered and in view"},
{"id": "5.4", "name": "Re-apply pressure further right", "segments": [[101, 1363]],
"pre": ["red_spot_visible(oozing_surface)"], "effects": ["on(gauze, oozing_surface)"],
"success_check": "gauze covers the spot that was still red"},
{"id": "5.5", "name": "Leave the gauze and free the grasper", "segments": [[101, 1405], [101, 1412]],
"pre": ["on(gauze, oozing_surface)"], "effects": ["out_of_view(#101)", "not held(gauze, #101)"],
"success_check": "gauze stays in place after release"},
{"id": "5.6", "name": "Probe the site", "segments": [[102, 1416], [102, 1445]],
"pre": ["visible(oozing_surface)"], "effects": [], "success_check": "rod tip reaches the spot"},
{"id": "5.7", "name": "Take the gauze again", "segments": [[3, 1470], [101, 1476]],
"pre": ["out_of_view(#101)"], "effects": ["held(gauze, #101)"], "success_check": "gauze moves with the grasper"},
],
"decisions": [
{"frame": 1346, "observation": "The gauze has been pressed on the area for about 3 s; what is under it cannot be seen.",
"assessment": "Pressure alone gives no feedback; the site has to be looked at.",
"decision": "Lift the gauze off the area.", "expected": "The oozing area comes into view.",
"outcome_seen": "Frames 1350-1360: the area is uncovered; a red spot is still visible near the lobules."},
{"frame": 1363, "observation": "A red spot is still visible after lifting, slightly right of where the gauze was.",
"assessment": "The first placement did not cover the spot well.",
"decision": "Put the gauze back and press it further to the right.", "expected": "The gauze covers the remaining spot.",
"outcome_seen": "Frames 1385-1404: the gauze is pushed up and to the right over the area."},
{"frame": 1405, "observation": "The gauze lies over the area.",
"assessment": "The grasper is not needed to keep the gauze there.",
"decision": "Release the gauze and withdraw the grasper.", "expected": "The gauze stays in place.",
"outcome_seen": "Frames 1410-1420: the grasper leaves; the gauze stays at the left of the field."},
{"frame": 1416, "observation": "The field is open; the rod is free.",
"assessment": "The spot can be checked by touch while the gauze rests.",
"decision": "Move the rod to the spot and touch it.", "expected": "The rod tip reaches the spot.",
"outcome_seen": "Frames 1425-1440: the rod tip is next to the red spot."},
]},
40: {
"goal": "Grasp and coagulate a tissue band at the base of a yellow lobule",
"initial_state": ["visible(tissue_band)", "visible(yellow_lobule)", "visible(purplish_tissue)", "on(gauze, left)"],
"subtasks": [
{"id": "40.1", "name": "Open and position the jaws around the band", "segments": [[101, 10000], [101, 10013], [101, 10051]],
"pre": ["visible(tissue_band)"], "effects": ["jaws_open(#101)", "between_jaws(tissue_band, #101)"],
"success_check": "the band sits between the open jaws"},
{"id": "40.2", "name": "Expose with the second instrument", "segments": [[102, 10126], [102, 10201]],
"pre": ["jaws_open(#101)"], "effects": ["visible(tissue_band)"],
"success_check": "tissue on the right is pushed away from the jaw tip"},
{"id": "40.3", "name": "Grasp the band", "segments": [[101, 10186], [101, 10201]],
"pre": ["between_jaws(tissue_band, #101)"], "effects": ["held(tissue_band, #101)"],
"success_check": "jaws closed with tissue between them"},
{"id": "40.4", "name": "Coagulate", "segments": [[101, 10225]],
"pre": ["held(tissue_band, #101)", "rod clear of the jaws"], "effects": ["energy_applied(tissue_band)"],
"success_check": "smoke and darkening at the jaw tip"},
],
"decisions": [
{"frame": 10051, "observation": "The open jaws are next to the band; the rod is low on the right.",
"assessment": "The band is not yet clearly between the jaws.",
"decision": "Hold the jaws open and make small adjustments.", "expected": "The band settles between the jaws.",
"outcome_seen": "Frames 10051-10185: the jaws stay open at the band with small movements."},
{"frame": 10126, "observation": "Tissue on the right lies against the jaw tip.",
"assessment": "It may be caught in the jaws or block the view.",
"decision": "Lift it away with the rod.", "expected": "The jaw tip and band are clear.",
"outcome_seen": "Frames 10126-10142: the rod lifts the tissue beside the jaws, then lowers again."},
{"frame": 10186, "observation": "The band lies between the open jaws.",
"assessment": "Ready to grasp.", "decision": "Close the jaws.", "expected": "The band is held.",
"outcome_seen": "Frames 10190-10200: the jaws close."},
{"frame": 10225, "observation": "The band is held; the rod has moved down and away from the tip.",
"assessment": "Energy can be applied without touching the rod.",
"decision": "Activate energy briefly.", "expected": "The grasped tissue coagulates.",
"outcome_seen": "Frames 10230-10249: smoke rises and the tissue at the tip darkens."},
]},
43: {
"goal": "Peel pale adherent material off a smooth pink surface",
"initial_state": ["held(pale_material, #102)", "visible(attachment_line)", "separated(pale_material, pink_surface, none)"],
"subtasks": [
{"id": "43.1", "name": "Put the material under tension", "segments": [[102, 10750]],
"pre": ["held(pale_material, #102)"], "effects": ["under_tension(pale_material)"],
"success_check": "strands between material and surface stretch"},
{"id": "43.2", "name": "Separate along the attachment", "segments": [[102, 10813], [101, 10785], [101, 10831]],
"pre": ["under_tension(pale_material)"], "effects": ["separated(pale_material, pink_surface, partial)"],
"success_check": "a sheet of material lifts off the surface"},
{"id": "43.3", "name": "Change the pulling direction", "segments": [[102, 10871]],
"pre": ["separated(pale_material, pink_surface, partial)"], "effects": ["under_tension(pale_material)"],
"success_check": "material pulled from a new angle"},
{"id": "43.4", "name": "Separate further", "segments": [[102, 10899], [101, 10901]],
"pre": ["under_tension(pale_material)"], "effects": ["separated(pale_material, pink_surface, partial)"],
"success_check": "more of the surface is uncovered"},
{"id": "43.5", "name": "Take hold again", "segments": [[102, 10940], [101, 10940]],
"pre": ["out_of_view(#102)"], "effects": ["held(pale_material, #102)"], "success_check": "jaws on the material"},
{"id": "43.6", "name": "Inspect", "segments": [["camera", 10975]],
"pre": ["out_of_view(#101)", "out_of_view(#102)"], "effects": [],
"success_check": "view held still on the peeled area"},
],
"decisions": [
{"frame": 10785, "observation": "The material is under tension; a line of attachment to the surface is visible.",
"assessment": "Traction alone tears unevenly; the attachment needs to be opened.",
"decision": "Push the rod tip into the attachment line.", "expected": "The material starts to come away along the line.",
"outcome_seen": "Frames 10800-10830: the rod tip is at the junction and the material starts to lift."},
{"frame": 10831, "observation": "A sheet of material is lifting while the grasper pulls left.",
"assessment": "The surface underneath moves with it.",
"decision": "Press the surface down with the rod (counter-traction).", "expected": "The material separates instead of dragging the surface.",
"outcome_seen": "Frames 10835-10850: a large sheet comes off to the left."},
{"frame": 10871, "observation": "The material is pulled as far left as it goes.",
"assessment": "Pulling further in the same direction gains nothing.",
"decision": "Bring the material back to the right to change the angle.", "expected": "A new part of the attachment comes under tension.",
"outcome_seen": "Frames 10875-10898: the grasper moves the material down and right."},
{"frame": 10975, "observation": "Both instruments have left the view.",
"assessment": "Time to see how much has come off.",
"decision": "Hold the camera still on the area.", "expected": "Residual material, if any, is visible.",
"outcome_seen": "Frames 10975-10999: a patch of pale material is still attached at the upper middle; removal is partial."},
{"frame": 10928, "observation": "The grasper leaves the view and comes back about 0.5 s later to take hold again.",
"assessment": "Whether the grip slipped cannot be seen.",
"decision": "Re-grasp the material.", "expected": "The material is held again.",
"outcome_seen": "Frames 10940-10954: the jaws are back on the material, then leave the view."},
]},
}
# ---- template-derived structure for the other clips
PREDICATES.update({"contact(I, X)": "the tip of instrument I touches X", "near(I, X)": "instrument I is next to X"})
RULES = { # verb: (preconditions, effects); I = acting instrument, X = target
"grasp": (["between_jaws(X, I)"], ["held(X, I)"]), "regrasp": (["held(X, I)"], ["held(X, I)"]),
"retract": (["held(X, I)"], ["under_tension(X)"]), "peel": (["held(X, I)", "under_tension(X)"], ["separated(X, surface, partial)"]),
"coagulate": (["held(X, I) or contact(I, X)"], ["energy_applied(X)"]), "press": ([], ["on(I, X)"]),
"release": (["held(X, I)"], ["not held(X, I)"]), "position": (["jaws_open(I)"], ["between_jaws(X, I)"]),
"open_jaws": ([], ["jaws_open(I)"]), "probe": ([], ["contact(I, X)"]),
"blunt_dissect": (["contact(I, X)"], ["separated(X, neighbour, partial)"]), "counter_traction": ([], ["under_tension(X)"]),
"transport": ([], ["near(I, X)"]), "approach": ([], ["near(I, X)"]), "enter": (["out_of_view(I)"], ["visible(I)"]),
"withdraw": ([], ["out_of_view(I)"]), "lift_off": (["on(I, X)"], ["visible(X)"]), "hold": ([], []), "sweep": ([], []),
"absent": ([], ["out_of_view(I)"])}
NAME = {"grasp": "Grasp", "retract": "Retract", "peel": "Peel", "coagulate": "Coagulate", "press": "Press on",
"probe": "Probe", "blunt_dissect": "Separate", "position": "Position the jaws at", "transport": "Move",
"release": "Release", "counter_traction": "Hold back", "lift_off": "Lift", "open_jaws": "Open the jaws at",
"hold": "Hold", "approach": "Approach", "regrasp": "Re-grasp"}
PRIORITY = ["coagulate", "peel", "blunt_dissect", "retract", "grasp", "press", "lift_off", "probe", "position",
"counter_traction", "release", "regrasp", "open_jaws", "transport", "approach", "hold"]
NAMES = {} # tissue id -> name for the clip being processed
def clean(t):
"""'attachment line between [202] and [201]' -> tag replaced by the tissue name; '[id]' after a name dropped."""
t = re.sub(r"(\S.*?)\s*\[(\d+)\]", lambda m: m.group(1) if m.group(1).strip().endswith(NAMES.get(m.group(2), "\0"))
else f"{m.group(1)} {NAMES.get(m.group(2), '')}", t)
return re.sub(r"\s+", " ", t).strip()
def fill(xs, inst, target):
tname = clean(target) if target not in ("-", "") else "the field"
return [x.replace("X", tname).replace("I", f"#{inst}") if x.startswith(("held", "between", "on(", "contact", "near", "visible(I"))
else x.replace("(X", f"({tname}").replace("I)", f"#{inst})").replace("(I", f"(#{inst}") for x in xs]
def template(c):
NAMES.clear()
NAMES.update(c["tissue_tracks"])
segs = c["segments"]
act = {}
for s in segs:
if s["verb"] not in ("absent", "withdraw", "enter", "sweep"):
act[s["track_id"]] = act.get(s["track_id"], 0) + s["end_frame"] - s["start_frame"] + 1
if not act:
return []
prim = max(act, key=act.get)
# a subtask is a run of the main instrument's segments on one target; it ends after a release,
# withdraw or absence, and a new one starts at an approach/entry once something has been done
runs, cur = [], None
MOVE, END = {"approach", "enter", "transport"}, {"release", "withdraw", "absent"}
for i, s in sorted(((i, s) for i, s in enumerate(segs) if s["track_id"] == prim), key=lambda x: x[1]["start_frame"]):
key = clean(s["target"]) if s["target"] not in ("-", "") else None
acted = cur and any(segs[j]["verb"] not in MOVE | END | {"sweep", "hold"} for j in cur["idx"])
same = cur and not cur["closed"] and (key is None or cur["key"] in (None, key)) and not (s["verb"] in MOVE and acted)
if same:
cur["idx"].append(i)
cur["key"] = cur["key"] or key
else:
cur = {"key": key, "idx": [i], "closed": False}
runs.append(cur)
if s["verb"] in END:
cur["closed"] = True
subs = []
for k, r in enumerate(runs):
ss = [segs[i] for i in r["idx"]]
if all(s["verb"] in ("absent", "withdraw", "enter", "sweep") for s in ss):
continue
main = min((s["verb"] for s in ss if s["verb"] in PRIORITY), key=PRIORITY.index, default="hold")
tgt = next((s["target"] for s in ss if s["target"] not in ("-", "")), "-")
f0, f1 = ss[0]["start_frame"], ss[-1]["end_frame"]
pre = fill(RULES[ss[0]["verb"]][0], prim, tgt)
eff = []
for s in ss:
for e in fill(RULES[s["verb"]][1], prim, s["target"] if s["target"] not in ("-", "") else tgt):
if e not in eff:
eff.append(e)
lasting = [e for e in eff if not e.startswith(("near(", "jaws_open(", "out_of_view(", "visible(#"))]
eff = lasting or eff # keep passing states only when nothing lasting happened
assist = [i for i, s in enumerate(segs) if s["track_id"] != prim and s["verb"] not in ("absent",)
and s["start_frame"] <= f1 and s["end_frame"] >= f0]
subs.append({"id": f"{c['clip_id']}.{len(subs) + 1}", "name": f"{NAME.get(main, main)} {clean(tgt) if tgt != '-' else ''}".strip(),
"primary_instrument": prim, "segment_index": r["idx"], "assisting_segment_index": assist,
"frames": [f0, f1], "pre": pre, "effects": eff,
"success_check": "; ".join(eff) if eff else "position maintained", "derived": "template"})
return subs
doc = {"description": "Task hierarchy, symbolic state and decision points for every clip. Clips 5, 40 and 43: "
"subtasks written by hand. Other clips: subtasks derived from the action segments of the clip's "
"main instrument with fixed verb rules (RULES), decision points written by hand from the video. "
"Observations and outcomes describe what is visible; assessments and decisions are a plausible "
"rationale, not the operator's recorded intent. AI draft, reviewed and approved by the contributing physician on 2026-11-17.",
"predicates": PREDICATES, "verb_rules": {k: {"pre": v[0], "effects": v[1]} for k, v in RULES.items()}, "clips": []}
for c in acts["clips"]:
if c["clip_id"] in T:
t = dict(T[c["clip_id"]])
segs = {(s["track_id"], s["start_frame"]): i for i, s in enumerate(c["segments"])}
for st in t["subtasks"]:
st["segment_index"] = [segs.get(tuple(k)) for k in st["segments"] if k[0] != "camera"]
assert None not in st["segment_index"], (c["clip_id"], st["id"])
st["frames"] = [min(k[1] for k in st["segments"]),
max(c["segments"][i]["end_frame"] for i in st["segment_index"]) if st["segment_index"] else c["end_frame"]]
st["derived"] = "hand"
doc["clips"].append({"clip_id": c["clip_id"], **t})
else:
doc["clips"].append({"clip_id": c["clip_id"], "goal": c["visible_goal"], "initial_state": [],
"subtasks": template(c), "decisions": c["decisions"]})
json.dump(doc, open(pack / "layer5_actions" / "task_graph.json", "w"), indent=1)
print("clips", len(doc["clips"]), "subtasks", sum(len(c["subtasks"]) for c in doc["clips"]),
"decisions", sum(len(c["decisions"]) for c in doc["clips"]))