feat(chexpert): U-MultiClass PNU abnormality guidance + abnormality-guided VQA
Browse files- chexpert_classifier.py: 14 binary heads → 14×3 softmax (negative/
positive/uncertain per pathology, META-CXR / CheXpert U-MultiClass).
Add format_pnu/buckets_to_pnu as the single source of truth for the
PNU 3-section prompt string (shared with the oracle builder so GT and
predicted prompts are byte-identical).
- mimic_cxr_builder.py: GT chexpert.csv → PNU string (1→pos, 0→neg,
-1→uncertain, blank/NaN→neg). VQA now carries the SAME PNU context
(abnormality-guided VQA, RaDialog-style). O(1) image lookup (was
O(N²)); index every image so report-less studies still serve VQA.
- Drop the obsolete uncertain_policy knob (U-MultiClass is the only
behaviour now) from builder, dataset.py, dataset_resolver, config.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- configs/train_config.yaml +5 -6
- data/dataset.py +14 -13
- data/mimic_cxr_builder.py +86 -84
- model/chexpert_classifier.py +97 -55
- model/{image_encoder.py → rad_dino.py} +0 -0
- utils/dataset_resolver.py +6 -7
|
@@ -59,15 +59,14 @@ data:
|
|
| 59 |
mimic_cxr_root: "/path/to/MIMIC-CXR"
|
| 60 |
instruct_json: "data/data_files/mimic_cxr_instruct_unified.json"
|
| 61 |
|
| 62 |
-
# RaDialog abnormality guidance
|
| 63 |
-
# read from this CSV and baked into the prompt as
|
| 64 |
-
# "
|
|
|
|
|
|
|
| 65 |
# *chexpert*.csv under mimic_cxr_root; if none is found, structured_findings
|
| 66 |
# is null and abnormality guidance is silently DISABLED (loud warning).
|
| 67 |
mimic_chexpert_csv: null
|
| 68 |
-
# How CheXpert -1.0 (uncertain) is mapped: "ignore" (only 1.0 positive,
|
| 69 |
-
# default, matches the classifier head) | "positive" (treat -1.0 as positive).
|
| 70 |
-
mimic_uncertain_policy: "ignore"
|
| 71 |
# Optional VQA pairs dir with {train,valid,test}.json. null → skip VQA.
|
| 72 |
mimic_vqa_root: null
|
| 73 |
# Auto-build the unified JSON (with CheXpert labels) when the cached
|
|
|
|
| 59 |
mimic_cxr_root: "/path/to/MIMIC-CXR"
|
| 60 |
instruct_json: "data/data_files/mimic_cxr_instruct_unified.json"
|
| 61 |
|
| 62 |
+
# RaDialog abnormality guidance (U-MultiClass / META-CXR): the 14 CheXpert
|
| 63 |
+
# labels (oracle / GT) are read from this CSV and baked into the prompt as
|
| 64 |
+
# the PNU 3-section string ("Positive Abnormalities: ... / Negative ... /
|
| 65 |
+
# Uncertain ..."). CSV value → class: 1→positive, 0→negative, -1→uncertain,
|
| 66 |
+
# blank/NaN→negative. If left null the builder auto-discovers any
|
| 67 |
# *chexpert*.csv under mimic_cxr_root; if none is found, structured_findings
|
| 68 |
# is null and abnormality guidance is silently DISABLED (loud warning).
|
| 69 |
mimic_chexpert_csv: null
|
|
|
|
|
|
|
|
|
|
| 70 |
# Optional VQA pairs dir with {train,valid,test}.json. null → skip VQA.
|
| 71 |
mimic_vqa_root: null
|
| 72 |
# Auto-build the unified JSON (with CheXpert labels) when the cached
|
|
@@ -26,7 +26,7 @@ from torch.utils.data import Dataset
|
|
| 26 |
from PIL import Image
|
| 27 |
|
| 28 |
from .prompt_templates import build_training_sample
|
| 29 |
-
from model.
|
| 30 |
|
| 31 |
|
| 32 |
TaskType = Literal["findings", "impression", "report", "vqa", "mixed"]
|
|
@@ -289,7 +289,6 @@ def build_instruct_json(
|
|
| 289 |
vqa_data_root: Optional[str] = None,
|
| 290 |
report_mode: str = "split",
|
| 291 |
image_mode: str = "all_views_split",
|
| 292 |
-
uncertain_policy: str = "ignore",
|
| 293 |
) -> str:
|
| 294 |
"""
|
| 295 |
Build the unified MIMIC-CXR instruction JSON.
|
|
@@ -297,25 +296,27 @@ def build_instruct_json(
|
|
| 297 |
Thin delegate to `data.mimic_cxr_builder.build_mimic_cxr_instruct_json`,
|
| 298 |
which walks the pre-split MIMIC layout (train/valid/test), parses
|
| 299 |
findings/impression from the report .txt files, and bakes the 14 CheXpert
|
| 300 |
-
labels (oracle, from `*chexpert*.csv`) into `structured_findings` as
|
| 301 |
-
|
| 302 |
-
setup. `report_mode` / `image_mode` mirror
|
|
|
|
| 303 |
|
| 304 |
Output entries match the shared schema, e.g.:
|
| 305 |
{"image_path": "train/p10/p10000032/s50414267/02aa804e.jpg",
|
| 306 |
"task": "findings", "target": "The lungs are clear...",
|
| 307 |
"question": null,
|
| 308 |
-
"structured_findings": "
|
|
|
|
|
|
|
| 309 |
"split": "train", "study_id": "s50414267",
|
| 310 |
"subject_id": "p10000032"}
|
| 311 |
"""
|
| 312 |
from .mimic_cxr_builder import build_mimic_cxr_instruct_json
|
| 313 |
return build_mimic_cxr_instruct_json(
|
| 314 |
-
mimic_root
|
| 315 |
-
output_path
|
| 316 |
-
chexpert_csv
|
| 317 |
-
vqa_root
|
| 318 |
-
report_mode
|
| 319 |
-
image_mode
|
| 320 |
-
uncertain_policy = uncertain_policy,
|
| 321 |
)
|
|
|
|
| 26 |
from PIL import Image
|
| 27 |
|
| 28 |
from .prompt_templates import build_training_sample
|
| 29 |
+
from model.rad_dino import BioViLTEncoder
|
| 30 |
|
| 31 |
|
| 32 |
TaskType = Literal["findings", "impression", "report", "vqa", "mixed"]
|
|
|
|
| 289 |
vqa_data_root: Optional[str] = None,
|
| 290 |
report_mode: str = "split",
|
| 291 |
image_mode: str = "all_views_split",
|
|
|
|
| 292 |
) -> str:
|
| 293 |
"""
|
| 294 |
Build the unified MIMIC-CXR instruction JSON.
|
|
|
|
| 296 |
Thin delegate to `data.mimic_cxr_builder.build_mimic_cxr_instruct_json`,
|
| 297 |
which walks the pre-split MIMIC layout (train/valid/test), parses
|
| 298 |
findings/impression from the report .txt files, and bakes the 14 CheXpert
|
| 299 |
+
labels (oracle, from `*chexpert*.csv`) into `structured_findings` as the
|
| 300 |
+
PNU 3-section string (U-MultiClass, META-CXR format) — the RaDialog
|
| 301 |
+
image + abnormality-guidance setup. `report_mode` / `image_mode` mirror
|
| 302 |
+
the IU builder.
|
| 303 |
|
| 304 |
Output entries match the shared schema, e.g.:
|
| 305 |
{"image_path": "train/p10/p10000032/s50414267/02aa804e.jpg",
|
| 306 |
"task": "findings", "target": "The lungs are clear...",
|
| 307 |
"question": null,
|
| 308 |
+
"structured_findings": "Positive Abnormalities: None\\n
|
| 309 |
+
Negative Abnormalities: No Finding, ...\\n
|
| 310 |
+
Uncertain Abnormalities: None",
|
| 311 |
"split": "train", "study_id": "s50414267",
|
| 312 |
"subject_id": "p10000032"}
|
| 313 |
"""
|
| 314 |
from .mimic_cxr_builder import build_mimic_cxr_instruct_json
|
| 315 |
return build_mimic_cxr_instruct_json(
|
| 316 |
+
mimic_root = mimic_cxr_root,
|
| 317 |
+
output_path = output_path,
|
| 318 |
+
chexpert_csv = chexpert_csv,
|
| 319 |
+
vqa_root = vqa_data_root,
|
| 320 |
+
report_mode = report_mode,
|
| 321 |
+
image_mode = image_mode,
|
|
|
|
| 322 |
)
|
|
@@ -15,25 +15,34 @@ NOT the raw PhysioNet tree):
|
|
| 15 |
└── test /pNN/...
|
| 16 |
{anywhere under mimic_root}/ *chexpert*.csv (optional, auto-discovered)
|
| 17 |
|
| 18 |
-
RaDialog-style abnormality guidance
|
| 19 |
-
-----------------------------------
|
| 20 |
The 14 CheXpert labels are read from `mimic-cxr-2.0.0-chexpert.csv`
|
| 21 |
(CheXbert run on the ground-truth reports) and baked into the prompt as
|
| 22 |
-
`structured_findings`:
|
| 23 |
|
| 24 |
-
|
| 25 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 26 |
|
| 27 |
This is the *oracle* setting — GT labels, no trained image classifier and
|
| 28 |
-
no model change. The
|
| 29 |
-
`
|
| 30 |
-
(
|
|
|
|
|
|
|
| 31 |
|
| 32 |
VQA
|
| 33 |
---
|
| 34 |
-
VQA pairs live in
|
| 35 |
-
|
| 36 |
-
|
|
|
|
|
|
|
|
|
|
| 37 |
"""
|
| 38 |
|
| 39 |
import argparse
|
|
@@ -50,23 +59,11 @@ from typing import Dict, List, Optional, Tuple
|
|
| 50 |
_FINDINGS_RE = re.compile(r"FINDINGS\s*:\s*(.*?)(?=\n\s*[A-Z ]{3,}\s*:|\Z)", re.S | re.I)
|
| 51 |
_IMPRESSION_RE = re.compile(r"IMPRESSION\s*:\s*(.*?)(?=\n\s*[A-Z ]{3,}\s*:|\Z)", re.S | re.I)
|
| 52 |
|
| 53 |
-
# 14
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
"Lung Opacity",
|
| 59 |
-
"Lung Lesion",
|
| 60 |
-
"Edema",
|
| 61 |
-
"Consolidation",
|
| 62 |
-
"Pneumonia",
|
| 63 |
-
"Atelectasis",
|
| 64 |
-
"Pneumothorax",
|
| 65 |
-
"Pleural Effusion",
|
| 66 |
-
"Pleural Other",
|
| 67 |
-
"Fracture",
|
| 68 |
-
"Support Devices",
|
| 69 |
-
]
|
| 70 |
|
| 71 |
|
| 72 |
def _clean(txt: str) -> str:
|
|
@@ -83,7 +80,7 @@ def _parse_report(txt_path: Path) -> Tuple[Optional[str], Optional[str]]:
|
|
| 83 |
)
|
| 84 |
|
| 85 |
|
| 86 |
-
# ─── CheXpert CSV →
|
| 87 |
|
| 88 |
def _discover_chexpert_csv(mimic_root: Path, explicit: Optional[str]) -> Optional[Path]:
|
| 89 |
if explicit:
|
|
@@ -97,17 +94,26 @@ def _discover_chexpert_csv(mimic_root: Path, explicit: Optional[str]) -> Optiona
|
|
| 97 |
return None
|
| 98 |
|
| 99 |
|
| 100 |
-
def _load_chexpert_map(
|
| 101 |
-
csv_path: Path,
|
| 102 |
-
uncertain_policy: str = "ignore", # "ignore" → only 1.0 positive | "positive" → -1.0 also positive
|
| 103 |
-
) -> Dict[Tuple[str, str], str]:
|
| 104 |
"""
|
| 105 |
-
Return {(subject_id, study_id):
|
| 106 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 107 |
"""
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 111 |
|
| 112 |
out: Dict[Tuple[str, str], str] = {}
|
| 113 |
with open(csv_path, newline="") as f:
|
|
@@ -121,25 +127,17 @@ def _load_chexpert_map(
|
|
| 121 |
f"{csv_path} missing subject_id/study_id columns "
|
| 122 |
f"(have: {reader.fieldnames})"
|
| 123 |
)
|
| 124 |
-
label_cols = [(name, col[name.lower()]) for name in
|
| 125 |
if name.lower() in col]
|
| 126 |
|
| 127 |
for row in reader:
|
| 128 |
subj = str(row[subj_c]).strip().lstrip("p").split(".")[0]
|
| 129 |
study = str(row[study_c]).strip().lstrip("s").split(".")[0]
|
| 130 |
-
|
| 131 |
-
name
|
| 132 |
-
|
| 133 |
-
|
| 134 |
-
|
| 135 |
-
# genuine positives (drop a redundant "No Finding" if any
|
| 136 |
-
# pathology is also positive).
|
| 137 |
-
real = [p for p in positives if p != "No Finding"]
|
| 138 |
-
if real:
|
| 139 |
-
txt = ", ".join(real)
|
| 140 |
-
else:
|
| 141 |
-
txt = "No Finding"
|
| 142 |
-
out[(subj, study)] = f"Predicted Findings: {txt}"
|
| 143 |
return out
|
| 144 |
|
| 145 |
|
|
@@ -152,18 +150,17 @@ def build_mimic_cxr_instruct_json(
|
|
| 152 |
vqa_root: Optional[str] = None,
|
| 153 |
report_mode: str = "split", # "split" | "merged" | "split_cascade"
|
| 154 |
image_mode: str = "all_views_split", # "all_views_split" | "frontal_only_split" | "multi_image_merged"
|
| 155 |
-
uncertain_policy: str = "ignore", # how CheXpert -1.0 (uncertain) is treated
|
| 156 |
) -> str:
|
| 157 |
"""
|
| 158 |
Build the unified MIMIC-CXR instruction JSON.
|
| 159 |
|
| 160 |
report_mode mirrors iu_xray_builder:
|
| 161 |
"split" → findings + impression samples; BOTH carry the CheXpert
|
| 162 |
-
|
| 163 |
-
|
| 164 |
"merged" → one task=report sample, target "Findings: ...\n\n
|
| 165 |
-
Impression: ...", carries the CheXpert string.
|
| 166 |
-
"split_cascade" → findings sample carries the CheXpert string; the
|
| 167 |
impression sample instead carries "Findings: <GT
|
| 168 |
findings>" as context (findings→impression). Same
|
| 169 |
convention as the IU builder.
|
|
@@ -201,9 +198,9 @@ def build_mimic_cxr_instruct_json(
|
|
| 201 |
# ── CheXpert labels ───────────────────────────────────────────────────
|
| 202 |
csv_path = _discover_chexpert_csv(mimic_root, chexpert_csv)
|
| 203 |
if csv_path is not None:
|
| 204 |
-
chexpert_map = _load_chexpert_map(csv_path
|
| 205 |
print(f"[mimic_cxr_builder] CheXpert CSV: {csv_path} "
|
| 206 |
-
f"({len(chexpert_map):,} studies,
|
| 207 |
else:
|
| 208 |
chexpert_map = {}
|
| 209 |
print("[mimic_cxr_builder] WARNING: no *chexpert*.csv found under "
|
|
@@ -213,23 +210,24 @@ def build_mimic_cxr_instruct_json(
|
|
| 213 |
|
| 214 |
# ── Pass 1: index studies ─────────────────────────────────────────────
|
| 215 |
samples: List[Dict] = []
|
| 216 |
-
|
|
|
|
|
|
|
| 217 |
n_studies = n_missing_report = n_no_chexpert = 0
|
| 218 |
skipped_merged_no_impression = skipped_cascade_no_findings = 0
|
| 219 |
|
| 220 |
def _structured_for(subj: str, study: str) -> Optional[str]:
|
| 221 |
return chexpert_map.get((subj.lstrip("p"), study.lstrip("s")))
|
| 222 |
|
| 223 |
-
def
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 224 |
"""Yield path_fields dicts honouring image_mode (same rules as IU)."""
|
| 225 |
-
imgs = sorted(study_dir.glob("*.jpg"))
|
| 226 |
-
if not imgs:
|
| 227 |
-
return
|
| 228 |
-
def _rel(img: Path) -> str:
|
| 229 |
-
return f"{split_sub}/{img.parent.parent.parent.name}/{subj}/{study}/{img.name}"
|
| 230 |
-
rels = [_rel(im) for im in imgs]
|
| 231 |
-
for r in rels:
|
| 232 |
-
image_index[r] = split_dirs[split_sub]
|
| 233 |
if image_mode == "all_views_split":
|
| 234 |
for r in rels:
|
| 235 |
yield {"image_path": r, "image_paths": None}
|
|
@@ -242,11 +240,15 @@ def build_mimic_cxr_instruct_json(
|
|
| 242 |
for p_dir in sorted(split_dir.glob("p*")):
|
| 243 |
for pat_dir in p_dir.glob("p*"):
|
| 244 |
for study_dir in pat_dir.glob("s*"):
|
| 245 |
-
|
| 246 |
-
|
|
|
|
| 247 |
continue
|
| 248 |
n_studies += 1
|
| 249 |
-
|
|
|
|
|
|
|
|
|
|
| 250 |
txts = list(study_dir.glob("*.txt"))
|
| 251 |
if not txts:
|
| 252 |
n_missing_report += 1
|
|
@@ -257,7 +259,7 @@ def build_mimic_cxr_instruct_json(
|
|
| 257 |
n_no_chexpert += 1
|
| 258 |
split_label = split_dirs[split_sub]
|
| 259 |
|
| 260 |
-
for path_fields in _image_groups(
|
| 261 |
base = {
|
| 262 |
**path_fields,
|
| 263 |
"question": None,
|
|
@@ -310,22 +312,26 @@ def build_mimic_cxr_instruct_json(
|
|
| 310 |
sub_rel = str(row["image_path"]).lstrip("/")
|
| 311 |
if sub_rel.startswith("files/"):
|
| 312 |
sub_rel = sub_rel[len("files/"):]
|
| 313 |
-
|
| 314 |
-
|
| 315 |
-
if hit is None:
|
| 316 |
n_vqa_dropped += 1
|
| 317 |
continue
|
| 318 |
ans = row.get("answer", [])
|
| 319 |
answer = (", ".join(map(str, ans)) if isinstance(ans, list)
|
| 320 |
else str(ans)) or "No."
|
|
|
|
|
|
|
| 321 |
samples.append({
|
| 322 |
-
"image_path":
|
| 323 |
"task": "vqa", "target": answer,
|
| 324 |
"question": row["question"],
|
| 325 |
-
|
|
|
|
|
|
|
|
|
|
| 326 |
"split": split_label,
|
| 327 |
-
"study_id":
|
| 328 |
-
"subject_id":
|
| 329 |
})
|
| 330 |
n_vqa += 1
|
| 331 |
|
|
@@ -372,9 +378,6 @@ def _parse_args():
|
|
| 372 |
choices=["split", "merged", "split_cascade"])
|
| 373 |
p.add_argument("--image_mode", default="all_views_split",
|
| 374 |
choices=["all_views_split", "frontal_only_split", "multi_image_merged"])
|
| 375 |
-
p.add_argument("--uncertain_policy", default="ignore",
|
| 376 |
-
choices=["ignore", "positive"],
|
| 377 |
-
help="CheXpert -1.0 (uncertain): ignore (default) or treat as positive.")
|
| 378 |
return p.parse_args()
|
| 379 |
|
| 380 |
|
|
@@ -387,5 +390,4 @@ if __name__ == "__main__":
|
|
| 387 |
vqa_root = a.vqa_root,
|
| 388 |
report_mode = a.report_mode,
|
| 389 |
image_mode = a.image_mode,
|
| 390 |
-
uncertain_policy = a.uncertain_policy,
|
| 391 |
)
|
|
|
|
| 15 |
└── test /pNN/...
|
| 16 |
{anywhere under mimic_root}/ *chexpert*.csv (optional, auto-discovered)
|
| 17 |
|
| 18 |
+
RaDialog-style abnormality guidance (U-MultiClass / META-CXR)
|
| 19 |
+
-------------------------------------------------------------
|
| 20 |
The 14 CheXpert labels are read from `mimic-cxr-2.0.0-chexpert.csv`
|
| 21 |
(CheXbert run on the ground-truth reports) and baked into the prompt as
|
| 22 |
+
`structured_findings` in the PNU 3-section format:
|
| 23 |
|
| 24 |
+
Positive Abnormalities: Cardiomegaly, Pleural Effusion
|
| 25 |
+
Negative Abnormalities: No Finding, Edema, ...
|
| 26 |
+
Uncertain Abnormalities: Atelectasis
|
| 27 |
+
|
| 28 |
+
CSV value → class: 1 → positive, 0 → negative, -1 → uncertain,
|
| 29 |
+
blank/NaN → negative (META-CXR convention: missing == negative).
|
| 30 |
|
| 31 |
This is the *oracle* setting — GT labels, no trained image classifier and
|
| 32 |
+
no model change. The string format is shared verbatim with
|
| 33 |
+
`model.chexpert_classifier.format_pnu`, so the learned-classifier path
|
| 34 |
+
(at inference) produces byte-identical prompts. The existing
|
| 35 |
+
`structured_findings` plumbing carries it through train (dataset.py) and
|
| 36 |
+
eval (evaluate.py) untouched.
|
| 37 |
|
| 38 |
VQA
|
| 39 |
---
|
| 40 |
+
VQA pairs live in 3 files {train,valid,test}.json (MIMIC-Ext-CXR-VQA);
|
| 41 |
+
attach them by passing `vqa_root`. Each row is one (image, question,
|
| 42 |
+
answer) sample — one image can yield many rows. VQA samples get the SAME
|
| 43 |
+
PNU CheXpert context as findings/impression (abnormality-guided VQA, à la
|
| 44 |
+
RaDialog), looked up by subject_id/study_id. Omit `vqa_root` to build
|
| 45 |
+
findings/impression only.
|
| 46 |
"""
|
| 47 |
|
| 48 |
import argparse
|
|
|
|
| 59 |
_FINDINGS_RE = re.compile(r"FINDINGS\s*:\s*(.*?)(?=\n\s*[A-Z ]{3,}\s*:|\Z)", re.S | re.I)
|
| 60 |
_IMPRESSION_RE = re.compile(r"IMPRESSION\s*:\s*(.*?)(?=\n\s*[A-Z ]{3,}\s*:|\Z)", re.S | re.I)
|
| 61 |
|
| 62 |
+
# The 14-label list, PNU string formatter and class indices live in
|
| 63 |
+
# model.chexpert_classifier — single source of truth shared with the learned
|
| 64 |
+
# classifier so GT-oracle and predicted prompts are byte-identical. Imported
|
| 65 |
+
# lazily inside _load_chexpert_map (it pulls the model package, which is
|
| 66 |
+
# always available in the train/eval env where JSON building runs).
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 67 |
|
| 68 |
|
| 69 |
def _clean(txt: str) -> str:
|
|
|
|
| 80 |
)
|
| 81 |
|
| 82 |
|
| 83 |
+
# ─── CheXpert CSV → PNU structured-findings string ──────────────────────────
|
| 84 |
|
| 85 |
def _discover_chexpert_csv(mimic_root: Path, explicit: Optional[str]) -> Optional[Path]:
|
| 86 |
if explicit:
|
|
|
|
| 94 |
return None
|
| 95 |
|
| 96 |
|
| 97 |
+
def _load_chexpert_map(csv_path: Path) -> Dict[Tuple[str, str], str]:
|
|
|
|
|
|
|
|
|
|
| 98 |
"""
|
| 99 |
+
Return {(subject_id, study_id): <PNU string>} where the ids are the bare
|
| 100 |
+
integers as strings (CSV stores them without the p/s prefix).
|
| 101 |
+
|
| 102 |
+
U-MultiClass mapping of each CheXpert cell:
|
| 103 |
+
1 / 1.0 → positive
|
| 104 |
+
0 / 0.0 → negative
|
| 105 |
+
-1 / -1.0 → uncertain
|
| 106 |
+
blank/NaN → negative (META-CXR convention: missing == negative)
|
| 107 |
"""
|
| 108 |
+
from model.chexpert_classifier import (
|
| 109 |
+
PATHOLOGIES, buckets_to_pnu,
|
| 110 |
+
CLASS_NEGATIVE, CLASS_POSITIVE, CLASS_UNCERTAIN,
|
| 111 |
+
)
|
| 112 |
+
val_to_cls = {
|
| 113 |
+
"1": CLASS_POSITIVE, "1.0": CLASS_POSITIVE,
|
| 114 |
+
"0": CLASS_NEGATIVE, "0.0": CLASS_NEGATIVE,
|
| 115 |
+
"-1": CLASS_UNCERTAIN, "-1.0": CLASS_UNCERTAIN,
|
| 116 |
+
}
|
| 117 |
|
| 118 |
out: Dict[Tuple[str, str], str] = {}
|
| 119 |
with open(csv_path, newline="") as f:
|
|
|
|
| 127 |
f"{csv_path} missing subject_id/study_id columns "
|
| 128 |
f"(have: {reader.fieldnames})"
|
| 129 |
)
|
| 130 |
+
label_cols = [(name, col[name.lower()]) for name in PATHOLOGIES
|
| 131 |
if name.lower() in col]
|
| 132 |
|
| 133 |
for row in reader:
|
| 134 |
subj = str(row[subj_c]).strip().lstrip("p").split(".")[0]
|
| 135 |
study = str(row[study_c]).strip().lstrip("s").split(".")[0]
|
| 136 |
+
mapping = {
|
| 137 |
+
name: val_to_cls.get(str(row.get(c, "")).strip(), CLASS_NEGATIVE)
|
| 138 |
+
for name, c in label_cols
|
| 139 |
+
}
|
| 140 |
+
out[(subj, study)] = buckets_to_pnu(mapping)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 141 |
return out
|
| 142 |
|
| 143 |
|
|
|
|
| 150 |
vqa_root: Optional[str] = None,
|
| 151 |
report_mode: str = "split", # "split" | "merged" | "split_cascade"
|
| 152 |
image_mode: str = "all_views_split", # "all_views_split" | "frontal_only_split" | "multi_image_merged"
|
|
|
|
| 153 |
) -> str:
|
| 154 |
"""
|
| 155 |
Build the unified MIMIC-CXR instruction JSON.
|
| 156 |
|
| 157 |
report_mode mirrors iu_xray_builder:
|
| 158 |
"split" → findings + impression samples; BOTH carry the CheXpert
|
| 159 |
+
PNU string in structured_findings (RaDialog: image +
|
| 160 |
+
14 labels → text).
|
| 161 |
"merged" → one task=report sample, target "Findings: ...\n\n
|
| 162 |
+
Impression: ...", carries the CheXpert PNU string.
|
| 163 |
+
"split_cascade" → findings sample carries the CheXpert PNU string; the
|
| 164 |
impression sample instead carries "Findings: <GT
|
| 165 |
findings>" as context (findings→impression). Same
|
| 166 |
convention as the IU builder.
|
|
|
|
| 198 |
# ── CheXpert labels ───────────────────────────────────────────────────
|
| 199 |
csv_path = _discover_chexpert_csv(mimic_root, chexpert_csv)
|
| 200 |
if csv_path is not None:
|
| 201 |
+
chexpert_map = _load_chexpert_map(csv_path)
|
| 202 |
print(f"[mimic_cxr_builder] CheXpert CSV: {csv_path} "
|
| 203 |
+
f"({len(chexpert_map):,} studies, PNU U-MultiClass)")
|
| 204 |
else:
|
| 205 |
chexpert_map = {}
|
| 206 |
print("[mimic_cxr_builder] WARNING: no *chexpert*.csv found under "
|
|
|
|
| 210 |
|
| 211 |
# ── Pass 1: index studies ─────────────────────────────────────────────
|
| 212 |
samples: List[Dict] = []
|
| 213 |
+
# sub_rel ("pXX/pXXXX/sYYYY/img.jpg") → full stored image_path
|
| 214 |
+
# ("{split}/pXX/pXXXX/sYYYY/img.jpg"). O(1) VQA lookup.
|
| 215 |
+
image_index: Dict[str, str] = {}
|
| 216 |
n_studies = n_missing_report = n_no_chexpert = 0
|
| 217 |
skipped_merged_no_impression = skipped_cascade_no_findings = 0
|
| 218 |
|
| 219 |
def _structured_for(subj: str, study: str) -> Optional[str]:
|
| 220 |
return chexpert_map.get((subj.lstrip("p"), study.lstrip("s")))
|
| 221 |
|
| 222 |
+
def _rels_for(study_dir: Path, split_sub: str, subj: str, study: str) -> List[str]:
|
| 223 |
+
"""Split-prefixed relative image paths for one study, sorted."""
|
| 224 |
+
return [
|
| 225 |
+
f"{split_sub}/{im.parent.parent.parent.name}/{subj}/{study}/{im.name}"
|
| 226 |
+
for im in sorted(study_dir.glob("*.jpg"))
|
| 227 |
+
]
|
| 228 |
+
|
| 229 |
+
def _image_groups(rels: List[str]):
|
| 230 |
"""Yield path_fields dicts honouring image_mode (same rules as IU)."""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 231 |
if image_mode == "all_views_split":
|
| 232 |
for r in rels:
|
| 233 |
yield {"image_path": r, "image_paths": None}
|
|
|
|
| 240 |
for p_dir in sorted(split_dir.glob("p*")):
|
| 241 |
for pat_dir in p_dir.glob("p*"):
|
| 242 |
for study_dir in pat_dir.glob("s*"):
|
| 243 |
+
subj, study = pat_dir.name, study_dir.name
|
| 244 |
+
rels = _rels_for(study_dir, split_sub, subj, study)
|
| 245 |
+
if not rels:
|
| 246 |
continue
|
| 247 |
n_studies += 1
|
| 248 |
+
# Index EVERY image up front — a VQA row may reference a
|
| 249 |
+
# study that has images but no findings/impression report.
|
| 250 |
+
for r in rels:
|
| 251 |
+
image_index[r.split("/", 1)[1]] = r
|
| 252 |
txts = list(study_dir.glob("*.txt"))
|
| 253 |
if not txts:
|
| 254 |
n_missing_report += 1
|
|
|
|
| 259 |
n_no_chexpert += 1
|
| 260 |
split_label = split_dirs[split_sub]
|
| 261 |
|
| 262 |
+
for path_fields in _image_groups(rels):
|
| 263 |
base = {
|
| 264 |
**path_fields,
|
| 265 |
"question": None,
|
|
|
|
| 312 |
sub_rel = str(row["image_path"]).lstrip("/")
|
| 313 |
if sub_rel.startswith("files/"):
|
| 314 |
sub_rel = sub_rel[len("files/"):]
|
| 315 |
+
full = image_index.get(sub_rel) # O(1)
|
| 316 |
+
if full is None:
|
|
|
|
| 317 |
n_vqa_dropped += 1
|
| 318 |
continue
|
| 319 |
ans = row.get("answer", [])
|
| 320 |
answer = (", ".join(map(str, ans)) if isinstance(ans, list)
|
| 321 |
else str(ans)) or "No."
|
| 322 |
+
subj = str(row.get("subject_id", ""))
|
| 323 |
+
study = str(row.get("study_id", ""))
|
| 324 |
samples.append({
|
| 325 |
+
"image_path": full, "image_paths": None,
|
| 326 |
"task": "vqa", "target": answer,
|
| 327 |
"question": row["question"],
|
| 328 |
+
# Abnormality-guided VQA (RaDialog): same PNU CheXpert
|
| 329 |
+
# context as findings/impression. None if no chexpert.csv
|
| 330 |
+
# (graceful — falls back to image + question only).
|
| 331 |
+
"structured_findings": _structured_for(subj, study),
|
| 332 |
"split": split_label,
|
| 333 |
+
"study_id": study,
|
| 334 |
+
"subject_id": subj,
|
| 335 |
})
|
| 336 |
n_vqa += 1
|
| 337 |
|
|
|
|
| 378 |
choices=["split", "merged", "split_cascade"])
|
| 379 |
p.add_argument("--image_mode", default="all_views_split",
|
| 380 |
choices=["all_views_split", "frontal_only_split", "multi_image_merged"])
|
|
|
|
|
|
|
|
|
|
| 381 |
return p.parse_args()
|
| 382 |
|
| 383 |
|
|
|
|
| 390 |
vqa_root = a.vqa_root,
|
| 391 |
report_mode = a.report_mode,
|
| 392 |
image_mode = a.image_mode,
|
|
|
|
| 393 |
)
|
|
@@ -1,21 +1,28 @@
|
|
| 1 |
"""
|
| 2 |
chexpert_classifier.py
|
| 3 |
----------------------
|
| 4 |
-
Multi-label CheXpert pathology classifier.
|
| 5 |
-
Trained separately on MIMIC-CXR with CheXbert labels.
|
| 6 |
|
| 7 |
-
|
| 8 |
-
|
| 9 |
-
|
| 10 |
|
| 11 |
-
|
| 12 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 13 |
"""
|
| 14 |
|
| 15 |
import torch
|
| 16 |
import torch.nn as nn
|
| 17 |
-
from
|
| 18 |
-
from typing import Optional, List, Dict
|
| 19 |
|
| 20 |
|
| 21 |
PATHOLOGIES = [
|
|
@@ -35,40 +42,77 @@ PATHOLOGIES = [
|
|
| 35 |
"Support Devices",
|
| 36 |
]
|
| 37 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 38 |
|
| 39 |
class CheXpertClassifier(nn.Module):
|
| 40 |
"""
|
| 41 |
-
|
| 42 |
|
| 43 |
-
|
| 44 |
-
|
| 45 |
|
| 46 |
Args:
|
| 47 |
-
input_dim:
|
| 48 |
-
num_classes: number of
|
| 49 |
-
|
| 50 |
-
checkpoint: path to trained weights (None = random init / not loaded)
|
| 51 |
"""
|
| 52 |
|
| 53 |
def __init__(
|
| 54 |
self,
|
| 55 |
-
input_dim:
|
| 56 |
num_classes: int = 14,
|
| 57 |
-
|
| 58 |
-
checkpoint: Optional[str] = None,
|
| 59 |
):
|
| 60 |
super().__init__()
|
| 61 |
|
| 62 |
self.num_classes = num_classes
|
| 63 |
-
self.
|
| 64 |
self.pathologies = PATHOLOGIES
|
| 65 |
|
| 66 |
-
#
|
| 67 |
self.classifier = nn.Sequential(
|
| 68 |
nn.Linear(input_dim, 256),
|
| 69 |
nn.ReLU(),
|
| 70 |
nn.Dropout(0.2),
|
| 71 |
-
nn.Linear(256, num_classes),
|
| 72 |
)
|
| 73 |
|
| 74 |
if checkpoint is not None:
|
|
@@ -82,46 +126,44 @@ class CheXpertClassifier(nn.Module):
|
|
| 82 |
def forward(self, global_features: torch.Tensor) -> torch.Tensor:
|
| 83 |
"""
|
| 84 |
Args:
|
| 85 |
-
global_features: (B, input_dim)
|
| 86 |
|
| 87 |
Returns:
|
| 88 |
-
logits: (B,
|
|
|
|
|
|
|
|
|
|
| 89 |
"""
|
| 90 |
-
|
|
|
|
| 91 |
|
| 92 |
@torch.no_grad()
|
| 93 |
def predict(self, global_features: torch.Tensor) -> List[Dict[str, str]]:
|
| 94 |
"""
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
Returns:
|
| 98 |
-
List of dicts like {"Pleural Effusion": "Positive", "Cardiomegaly": "Negative", ...}
|
| 99 |
-
"""
|
| 100 |
-
logits = self.forward(global_features) # (B, 14)
|
| 101 |
-
probs = torch.sigmoid(logits) # (B, 14)
|
| 102 |
-
preds = (probs > self.threshold).cpu() # (B, 14) bool
|
| 103 |
-
|
| 104 |
-
results = []
|
| 105 |
-
for i in range(preds.size(0)):
|
| 106 |
-
finding = {}
|
| 107 |
-
for j, name in enumerate(self.pathologies):
|
| 108 |
-
finding[name] = "Positive" if preds[i, j].item() else "Negative"
|
| 109 |
-
results.append(finding)
|
| 110 |
-
return results
|
| 111 |
-
|
| 112 |
-
def findings_to_text(self, findings: Dict[str, str]) -> str:
|
| 113 |
"""
|
| 114 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 115 |
|
| 116 |
-
|
| 117 |
-
|
| 118 |
"""
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
|
| 123 |
-
|
| 124 |
-
|
| 125 |
-
|
| 126 |
-
|
| 127 |
-
|
|
|
|
|
|
|
|
|
| 1 |
"""
|
| 2 |
chexpert_classifier.py
|
| 3 |
----------------------
|
| 4 |
+
Multi-label, multi-CLASS CheXpert pathology classifier (U-MultiClass).
|
|
|
|
| 5 |
|
| 6 |
+
Each of the 14 pathologies is predicted as one of THREE classes —
|
| 7 |
+
negative / positive / uncertain — via a per-pathology softmax, mirroring
|
| 8 |
+
META-CXR's MHCAC head and the CheXpert "U-MultiClass" uncertainty policy.
|
| 9 |
|
| 10 |
+
The structured findings injected into the LLM prompt use the PNU
|
| 11 |
+
(Positive / Negative / Uncertain) 3-section format. `format_pnu()` is the
|
| 12 |
+
single source of truth for that string so the oracle path
|
| 13 |
+
(data/mimic_cxr_builder.py, GT from chexpert.csv) and the learned path
|
| 14 |
+
(this classifier at inference) produce byte-identical prompts.
|
| 15 |
+
|
| 16 |
+
Trained separately (Stage 0) on MIMIC-CXR CheXbert labels; frozen during
|
| 17 |
+
Stage 1 / Stage 2 of the main VLM.
|
| 18 |
+
|
| 19 |
+
Reference: RaDialog (Pellegrini et al., 2023) for the prompt-conditioning
|
| 20 |
+
idea; META-CXR (Edirisinghe et al., 2025) for the explicit uncertain class.
|
| 21 |
"""
|
| 22 |
|
| 23 |
import torch
|
| 24 |
import torch.nn as nn
|
| 25 |
+
from typing import Optional, List, Dict, Sequence
|
|
|
|
| 26 |
|
| 27 |
|
| 28 |
PATHOLOGIES = [
|
|
|
|
| 42 |
"Support Devices",
|
| 43 |
]
|
| 44 |
|
| 45 |
+
# Per-pathology class indices (softmax dim order). Keep this stable: the
|
| 46 |
+
# trained checkpoint and the GT-label mapping in mimic_cxr_builder.py both
|
| 47 |
+
# rely on it.
|
| 48 |
+
CLASS_NEGATIVE = 0
|
| 49 |
+
CLASS_POSITIVE = 1
|
| 50 |
+
CLASS_UNCERTAIN = 2
|
| 51 |
+
NUM_STATES = 3
|
| 52 |
+
CLASS_NAMES = {CLASS_NEGATIVE: "negative",
|
| 53 |
+
CLASS_POSITIVE: "positive",
|
| 54 |
+
CLASS_UNCERTAIN: "uncertain"}
|
| 55 |
+
|
| 56 |
+
|
| 57 |
+
def format_pnu(positive: Sequence[str],
|
| 58 |
+
negative: Sequence[str],
|
| 59 |
+
uncertain: Sequence[str]) -> str:
|
| 60 |
+
"""
|
| 61 |
+
Build the PNU structured-findings string (META-CXR prompt format).
|
| 62 |
+
|
| 63 |
+
Positive Abnormalities: Cardiomegaly, Pleural Effusion
|
| 64 |
+
Negative Abnormalities: No Finding, Edema, ...
|
| 65 |
+
Uncertain Abnormalities: Atelectasis
|
| 66 |
+
|
| 67 |
+
Empty sections render as "None" so the three lines are always present
|
| 68 |
+
(the LLM sees a fixed structure regardless of the case).
|
| 69 |
+
"""
|
| 70 |
+
def _fmt(xs: Sequence[str]) -> str:
|
| 71 |
+
return ", ".join(xs) if xs else "None"
|
| 72 |
+
return (f"Positive Abnormalities: {_fmt(positive)}\n"
|
| 73 |
+
f"Negative Abnormalities: {_fmt(negative)}\n"
|
| 74 |
+
f"Uncertain Abnormalities: {_fmt(uncertain)}")
|
| 75 |
+
|
| 76 |
+
|
| 77 |
+
def buckets_to_pnu(class_by_pathology: Dict[str, int]) -> str:
|
| 78 |
+
"""Group a {pathology: class_idx} dict into the PNU string."""
|
| 79 |
+
pos = [p for p, c in class_by_pathology.items() if c == CLASS_POSITIVE]
|
| 80 |
+
neg = [p for p, c in class_by_pathology.items() if c == CLASS_NEGATIVE]
|
| 81 |
+
unc = [p for p, c in class_by_pathology.items() if c == CLASS_UNCERTAIN]
|
| 82 |
+
return format_pnu(pos, neg, unc)
|
| 83 |
+
|
| 84 |
|
| 85 |
class CheXpertClassifier(nn.Module):
|
| 86 |
"""
|
| 87 |
+
Multi-label, 3-class-per-label classifier on BioViL-T global embeddings.
|
| 88 |
|
| 89 |
+
Output logits have shape (B, 14, 3); a per-pathology softmax/argmax
|
| 90 |
+
yields negative / positive / uncertain.
|
| 91 |
|
| 92 |
Args:
|
| 93 |
+
input_dim: global CXR embedding dim
|
| 94 |
+
num_classes: number of pathologies (14)
|
| 95 |
+
checkpoint: trained weights (None = not loaded)
|
|
|
|
| 96 |
"""
|
| 97 |
|
| 98 |
def __init__(
|
| 99 |
self,
|
| 100 |
+
input_dim: int = 512,
|
| 101 |
num_classes: int = 14,
|
| 102 |
+
checkpoint: Optional[str] = None,
|
|
|
|
| 103 |
):
|
| 104 |
super().__init__()
|
| 105 |
|
| 106 |
self.num_classes = num_classes
|
| 107 |
+
self.num_states = NUM_STATES
|
| 108 |
self.pathologies = PATHOLOGIES
|
| 109 |
|
| 110 |
+
# MLP head → num_classes * 3 logits, reshaped to (B, num_classes, 3)
|
| 111 |
self.classifier = nn.Sequential(
|
| 112 |
nn.Linear(input_dim, 256),
|
| 113 |
nn.ReLU(),
|
| 114 |
nn.Dropout(0.2),
|
| 115 |
+
nn.Linear(256, num_classes * NUM_STATES),
|
| 116 |
)
|
| 117 |
|
| 118 |
if checkpoint is not None:
|
|
|
|
| 126 |
def forward(self, global_features: torch.Tensor) -> torch.Tensor:
|
| 127 |
"""
|
| 128 |
Args:
|
| 129 |
+
global_features: (B, input_dim)
|
| 130 |
|
| 131 |
Returns:
|
| 132 |
+
logits: (B, num_classes, 3) — softmax over the last dim gives
|
| 133 |
+
P(negative), P(positive), P(uncertain) per pathology.
|
| 134 |
+
Train with cross-entropy over the last dim (the natural
|
| 135 |
+
U-MultiClass objective).
|
| 136 |
"""
|
| 137 |
+
flat = self.classifier(global_features) # (B, 14*3)
|
| 138 |
+
return flat.view(-1, self.num_classes, NUM_STATES) # (B, 14, 3)
|
| 139 |
|
| 140 |
@torch.no_grad()
|
| 141 |
def predict(self, global_features: torch.Tensor) -> List[Dict[str, str]]:
|
| 142 |
"""
|
| 143 |
+
Returns a list (per sample) of {pathology: "negative"|"positive"|
|
| 144 |
+
"uncertain"} using argmax over the 3-state softmax.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 145 |
"""
|
| 146 |
+
logits = self.forward(global_features) # (B, 14, 3)
|
| 147 |
+
cls = logits.argmax(dim=-1).cpu() # (B, 14)
|
| 148 |
+
out: List[Dict[str, str]] = []
|
| 149 |
+
for i in range(cls.size(0)):
|
| 150 |
+
out.append({
|
| 151 |
+
name: CLASS_NAMES[int(cls[i, j].item())]
|
| 152 |
+
for j, name in enumerate(self.pathologies)
|
| 153 |
+
})
|
| 154 |
+
return out
|
| 155 |
|
| 156 |
+
@torch.no_grad()
|
| 157 |
+
def findings_to_text(self, global_features: torch.Tensor) -> List[str]:
|
| 158 |
"""
|
| 159 |
+
Per-sample PNU structured-findings string, identical in format to the
|
| 160 |
+
GT oracle path (data/mimic_cxr_builder.py). One string per sample.
|
| 161 |
+
"""
|
| 162 |
+
logits = self.forward(global_features) # (B, 14, 3)
|
| 163 |
+
cls = logits.argmax(dim=-1).cpu() # (B, 14)
|
| 164 |
+
texts: List[str] = []
|
| 165 |
+
for i in range(cls.size(0)):
|
| 166 |
+
mapping = {name: int(cls[i, j].item())
|
| 167 |
+
for j, name in enumerate(self.pathologies)}
|
| 168 |
+
texts.append(buckets_to_pnu(mapping))
|
| 169 |
+
return texts
|
|
File without changes
|
|
@@ -223,13 +223,12 @@ def _ensure_mimic_json_exists(data_cfg,
|
|
| 223 |
print(f"[dataset_resolver] MIMIC JSON not found → auto-building "
|
| 224 |
f"(report_mode={report_mode}, image_mode={image_mode}) …")
|
| 225 |
build_mimic_cxr_instruct_json(
|
| 226 |
-
mimic_root
|
| 227 |
-
output_path
|
| 228 |
-
chexpert_csv
|
| 229 |
-
vqa_root
|
| 230 |
-
report_mode
|
| 231 |
-
image_mode
|
| 232 |
-
uncertain_policy = str(_get(data_cfg, "mimic_uncertain_policy", "ignore")),
|
| 233 |
)
|
| 234 |
return str(out)
|
| 235 |
|
|
|
|
| 223 |
print(f"[dataset_resolver] MIMIC JSON not found → auto-building "
|
| 224 |
f"(report_mode={report_mode}, image_mode={image_mode}) …")
|
| 225 |
build_mimic_cxr_instruct_json(
|
| 226 |
+
mimic_root = str(_get(data_cfg, "mimic_cxr_root")),
|
| 227 |
+
output_path = str(out),
|
| 228 |
+
chexpert_csv = _get(data_cfg, "mimic_chexpert_csv"),
|
| 229 |
+
vqa_root = _get(data_cfg, "mimic_vqa_root"),
|
| 230 |
+
report_mode = report_mode,
|
| 231 |
+
image_mode = image_mode,
|
|
|
|
| 232 |
)
|
| 233 |
return str(out)
|
| 234 |
|