Fix loader: correct CADAD signature, bundle sentence-segmentation collate + concept vocab, decode_greedy path verified end-to-end

chest2err.py

@@ -2,35 +2,38 @@
 
 Usage:
     from chest2err import chest2err_score, chest2err_detail
+    score  = chest2err_score(ref, cand)         # float in (0, 1]
+    detail = chest2err_detail(ref, cand)        # full breakdown
 
-    score = chest2err_score(ref_report, candidate_report)         # float in (0, 1]
-    detail = chest2err_detail(ref_report, candidate_report)        # full breakdown
-
-The bundle ships the merged backbone + decoder weights and the Qwen3-architecture
-config, so no extra weights are downloaded at inference time. The backbone class
-itself is loaded from the `transformers` package.
+The bundle ships the merged backbone weights, the decoder weights, the
+tokenizer, and the concept vocabulary. No additional downloads occur at
+inference; the Qwen3-architecture backbone class is taken from the
+`transformers` package and instantiated from the bundled `config.json`.
 """
 from __future__ import annotations
 
 import json
-import os
-import re
 import math
+import os
 from pathlib import Path
-from typing import Any, Dict, List, Optional, Tuple
+from typing import Any, Dict, List, Optional
 
 import torch
-import torch.nn.functional as F
 from transformers import AutoModel, AutoTokenizer
 from safetensors.torch import load_file
 
-# Import the decoder module that ships in the same directory.
+# Sibling files in this package
 from chest2err_modeling import CADAD
-
-# ---------------------------------------------------------------------------
+from chest2err_collate import encode_pair_for_decoder, collate_decoder_batch
 
 PACKAGE_DIR = Path(__file__).resolve().parent
 
+CAT_NAMES = {0: "EOS", 1: "false_prediction", 2: "omission", 3: "location",
+             4: "severity", 5: "comparison"}
+ANAT_NAMES = {0: "Lung & Airways", 1: "Cardiovascular", 2: "Mediastinum & Hila",
+              3: "Upper Abdomen", 4: "Pleura", 5: "Bones / Spine", 6: "Chest Wall",
+              7: "Lower Neck", 8: "Others"}
+
 
 def _load_config() -> Dict[str, Any]:
     with open(PACKAGE_DIR / "chest2err_config.json") as f:
@@ -40,114 +43,106 @@ def _load_config() -> Dict[str, Any]:
 class Chest2Err:
     """Loads the merged backbone + decoder once, then scores pairs."""
 
-    def __init__(self, device: str = "cuda" if torch.cuda.is_available() else "cpu"):
+    def __init__(self,
+                 device: str = "cuda" if torch.cuda.is_available() else "cpu",
+                 attn_implementation: Optional[str] = None):
         cfg = _load_config()
         self.cfg = cfg
         self.device = device
         self.max_length = cfg["max_length"]
-        self.template = cfg["input_template"]
 
-        # Backbone: load the chest2vec_0.6b architecture from the bundled config + weights.
-        # No HuggingFace download — the safetensors and config.json are local to this package.
-        self.tokenizer = AutoTokenizer.from_pretrained(str(PACKAGE_DIR))
-        self.backbone = AutoModel.from_pretrained(
-            str(PACKAGE_DIR),
-            torch_dtype=torch.bfloat16,
-        ).to(device).eval()
+        # Concept vocab (size determines decoder output head dim)
+        with open(PACKAGE_DIR / "concept2id.json") as f:
+            self.concept2id: Dict[str, int] = json.load(f)
+        self.n_concept = len(self.concept2id)
+        self.id2concept = {v: k for k, v in self.concept2id.items()}
 
-        # Decoder + null embeddings + heads.
-        decoder_state = load_file(str(PACKAGE_DIR / "decoder.safetensors"))
-        n_concepts = decoder_state["concept_head.weight"].shape[0] if "concept_head.weight" in decoder_state else 1
-        self.decoder = CADAD(
-            hidden=cfg["hidden_size"],
-            n_cat=cfg["n_cat"] + 1,           # +1 for EOS at index 0
+        # Tokenizer + backbone load from bundled files only.
+        self.tokenizer = AutoTokenizer.from_pretrained(str(PACKAGE_DIR))
+        kw = {"torch_dtype": torch.bfloat16}
+        if attn_implementation:
+            kw["attn_implementation"] = attn_implementation
+        backbone = AutoModel.from_pretrained(str(PACKAGE_DIR), **kw)
+
+        # CADAD wraps the backbone + decoder. Construct, then load merged backbone
+        # weights + decoder weights.
+        self.model = CADAD(
+            backbone=backbone,
+            hidden_size=cfg["hidden_size"],
+            n_cat=cfg["n_cat"],
             n_anat=cfg["n_anat"],
-            n_concepts=n_concepts,
+            n_concept=self.n_concept,
+            n_severity=2,
             decoder_layers=cfg["decoder_layers"],
             decoder_heads=cfg["decoder_heads"],
             decoder_ff=cfg["decoder_ff"],
-            decoder_dropout=cfg["decoder_dropout"],
+            dropout=cfg["decoder_dropout"],
             max_decode_steps=cfg["max_decode_steps"],
         )
-        self.decoder.load_state_dict(decoder_state, strict=False)
-        self.decoder = self.decoder.to(device).to(torch.bfloat16).eval()
-
-    # ----------------------- input prep ------------------------- #
-
-    @staticmethod
-    def _split_sentences(text: str) -> List[str]:
-        """Light sentence splitter. Section headers and bullet lines count as boundaries too."""
-        # Split on . ! ? and section headers like [Lungs] or "Lungs:"
-        chunks = re.split(r"(?<=[.!?])\s+|\n+", text or "")
-        sents = [c.strip().lstrip("- ").strip() for c in chunks]
-        return [s for s in sents if s]
-
-    def _encode_pair(self, ref: str, cand: str) -> Dict[str, torch.Tensor]:
-        ref_sents = self._split_sentences(ref)
-        cand_sents = self._split_sentences(cand)
-        text = self.template.format(reference_report=ref, candidate_report=cand)
-        enc = self.tokenizer(
-            text,
-            max_length=self.max_length,
-            truncation=True,
-            padding=False,
-            return_tensors="pt",
-            add_special_tokens=False,
-        )
-        # NB: a production-grade encoder also produces seg_token_mask aligning each
-        # sentence to its token span. The CADAD decoder consumes per-sentence
-        # mean-pooled vectors; this helper exposes the API surface.
-        return {
-            "input_ids": enc["input_ids"].to(self.device),
-            "attention_mask": enc["attention_mask"].to(self.device),
-            "ref_sentences": ref_sents,
-            "cand_sentences": cand_sents,
-        }
 
-    # ----------------------- public API ------------------------- #
+        # The backbone weights were already loaded by AutoModel.from_pretrained.
+        # Now layer the decoder weights on top.
+        decoder_state = load_file(str(PACKAGE_DIR / "decoder.safetensors"))
+        missing, unexpected = self.model.load_state_dict(decoder_state, strict=False)
+        # Expected: many `backbone.*` keys are "missing" from decoder_state
+        # (they came from model.safetensors via from_pretrained). That's fine.
+
+        self.model = self.model.to(device).eval()
 
     @torch.inference_mode()
     def score(self, ref: str, cand: str) -> float:
-        """chest2err-score ∈ (0, 1]. Higher = better."""
-        detail = self.detail(ref, cand)
-        return detail["score"]
+        return self.detail(ref, cand)["score"]
 
     @torch.inference_mode()
     def detail(self, ref: str, cand: str) -> Dict[str, Any]:
-        """Full breakdown: score, K_total, per-error tuples, per-category and per-anatomy counts."""
-        enc = self._encode_pair(ref, cand)
-        out = self.backbone(
-            input_ids=enc["input_ids"],
-            attention_mask=enc["attention_mask"],
-            use_cache=False,
-        )
-        h = out.last_hidden_state
-        tuples = self.decoder.generate(
-            h=h,
-            attention_mask=enc["attention_mask"],
-            ref_sentences=enc["ref_sentences"],
-            cand_sentences=enc["cand_sentences"],
+        item = encode_pair_for_decoder(
+            self.tokenizer, ref, cand, max_length=self.max_length,
         )
-        K_total = len(tuples)
+        batch = collate_decoder_batch([item],
+                                       pad_token_id=self.tokenizer.pad_token_id or 0)
+        batch = {k: v.to(self.device) for k, v in batch.items()}
+
+        with torch.autocast(
+            device_type="cuda" if str(self.device).startswith("cuda") else "cpu",
+            dtype=torch.bfloat16,
+        ):
+            seqs = self.model.decode_greedy(
+                batch["input_ids"],
+                batch["attention_mask"],
+                batch["ref_seg_token_mask"],
+                batch["cand_seg_token_mask"],
+            )
+        seq = seqs[0]
+        K_total = len(seq)
         score = math.exp(-K_total)
+
         cat_counts = [0] * self.cfg["n_cat"]
         anat_counts = [0] * self.cfg["n_anat"]
-        for t in tuples:
-            if 1 <= t["cat"] <= self.cfg["n_cat"]:
-                cat_counts[t["cat"] - 1] += 1
-            if 0 <= t["anat"] < self.cfg["n_anat"]:
-                anat_counts[t["anat"]] += 1
+        tuples_out: List[Dict[str, Any]] = []
+        for t in seq:
+            c = int(t.get("cat", 0))
+            a = int(t.get("anat", 0))
+            if 1 <= c <= self.cfg["n_cat"]:
+                cat_counts[c - 1] += 1
+            if 0 <= a < self.cfg["n_anat"]:
+                anat_counts[a] += 1
+            tuples_out.append({
+                "cat": c, "cat_name": CAT_NAMES.get(c, str(c)),
+                "anat": a, "anat_name": ANAT_NAMES.get(a, str(a)),
+                "concept_id": int(t.get("concept_id", 0)),
+                "concept": self.id2concept.get(int(t.get("concept_id", 0)), "<UNK>"),
+                "ref_seg_idx": int(t.get("ref_seg_idx", -1)),
+                "cand_seg_idx": int(t.get("cand_seg_idx", -1)),
+            })
         return {
             "score": score,
             "K_total": K_total,
-            "tuples": tuples,
+            "tuples": tuples_out,
             "category_counts": cat_counts,
             "anatomy_counts": anat_counts,
         }
 
 
-# ----------------------- module-level convenience ----------------------- #
-
 _INSTANCE: Optional[Chest2Err] = None
 
 

chest2err_collate.py

@@ -0,0 +1,105 @@
+"""Sentence segmentation + collate for chest2err inference.
+
+Stripped-down version of the in-tree training collate: keeps only what's needed
+to run greedy decoding on one (ref, cand) pair.
+"""
+from __future__ import annotations
+
+import re
+from typing import Any, Dict, List, Tuple
+
+import torch
+
+
+_SECTION_HEADER_RE = re.compile(r"\[[^\]]+\]")
+_BULLET_RE = re.compile(r"\n\s*[-*•]\s+")
+_SENT_BOUNDARY_RE = re.compile(r"(?<=\.)\s+(?=[A-Z])")
+
+
+def segment_report(text: str, min_len: int = 3) -> List[str]:
+    if not text or not text.strip():
+        return []
+    t = _SECTION_HEADER_RE.sub("\n", text)
+    t = _BULLET_RE.sub("\n", t)
+    out: List[str] = []
+    for line in t.split("\n"):
+        line = line.strip()
+        if len(line) < min_len:
+            continue
+        for sent in _SENT_BOUNDARY_RE.split(line):
+            sent = sent.strip()
+            if len(sent) >= min_len:
+                out.append(sent)
+    return out
+
+
+def encode_pair_for_decoder(
+    tokenizer,
+    ref_text: str,
+    cand_text: str,
+    max_length: int = 1280,
+    ref_marker: str = "[REF]",
+    cand_marker: str = "[PRED]",
+) -> Dict[str, Any]:
+    ref_segs = segment_report(ref_text)
+    cand_segs = segment_report(cand_text)
+
+    ref_marker_ids = tokenizer.encode(ref_marker + " ", add_special_tokens=False)
+    cand_marker_ids = tokenizer.encode(" " + cand_marker + " ", add_special_tokens=False)
+    ref_seg_token_ids = [tokenizer.encode(s + " ", add_special_tokens=False) for s in ref_segs]
+    cand_seg_token_ids = [tokenizer.encode(s + " ", add_special_tokens=False) for s in cand_segs]
+
+    def _total_len(rs, cs):
+        return (len(ref_marker_ids) + sum(len(x) for x in rs)
+                + len(cand_marker_ids) + sum(len(x) for x in cs))
+
+    while _total_len(ref_seg_token_ids, cand_seg_token_ids) > max_length and cand_seg_token_ids:
+        cand_seg_token_ids.pop(); cand_segs = cand_segs[:-1]
+    while _total_len(ref_seg_token_ids, cand_seg_token_ids) > max_length and ref_seg_token_ids:
+        ref_seg_token_ids.pop(); ref_segs = ref_segs[:-1]
+
+    input_ids: List[int] = []
+    input_ids.extend(ref_marker_ids)
+    ref_ranges: List[Tuple[int, int]] = []
+    for ids in ref_seg_token_ids:
+        s = len(input_ids); input_ids.extend(ids); ref_ranges.append((s, len(input_ids)))
+    input_ids.extend(cand_marker_ids)
+    cand_ranges: List[Tuple[int, int]] = []
+    for ids in cand_seg_token_ids:
+        s = len(input_ids); input_ids.extend(ids); cand_ranges.append((s, len(input_ids)))
+
+    return {
+        "input_ids": input_ids,
+        "ref_seg_ranges": ref_ranges,
+        "cand_seg_ranges": cand_ranges,
+        "ref_segs": ref_segs,
+        "cand_segs": cand_segs,
+    }
+
+
+def collate_decoder_batch(items: List[Dict[str, Any]], pad_token_id: int = 0) -> Dict[str, torch.Tensor]:
+    T = max(len(it["input_ids"]) for it in items)
+    Sr = max(max(len(it["ref_seg_ranges"]), 1) for it in items)
+    Sc = max(max(len(it["cand_seg_ranges"]), 1) for it in items)
+    B = len(items)
+
+    input_ids = torch.full((B, T), pad_token_id, dtype=torch.long)
+    attention_mask = torch.zeros((B, T), dtype=torch.long)
+    ref_seg_token_mask = torch.zeros((B, Sr, T), dtype=torch.bool)
+    cand_seg_token_mask = torch.zeros((B, Sc, T), dtype=torch.bool)
+
+    for b, it in enumerate(items):
+        ids = it["input_ids"]; L = len(ids)
+        input_ids[b, :L] = torch.tensor(ids, dtype=torch.long)
+        attention_mask[b, :L] = 1
+        for s, (a, e) in enumerate(it["ref_seg_ranges"]):
+            ref_seg_token_mask[b, s, a:e] = True
+        for s, (a, e) in enumerate(it["cand_seg_ranges"]):
+            cand_seg_token_mask[b, s, a:e] = True
+
+    return {
+        "input_ids": input_ids,
+        "attention_mask": attention_mask,
+        "ref_seg_token_mask": ref_seg_token_mask,
+        "cand_seg_token_mask": cand_seg_token_mask,
+    }

concept2id.json

@@ -0,0 +1,387 @@
+{
+  "Abdominal aortic aneurysm": 0,
+  "Abdominal aortic aneurysm (partially imaged)": 1,
+  "Abdominal aortic calcification / atherosclerosis": 2,
+  "Abdominal aortic calcification / atherosclerosis (partially imaged)": 3,
+  "Abdominal lymphadenopathy": 4,
+  "Aberrant right subclavian artery": 5,
+  "Accessory hemiazygos vein": 6,
+  "Accessory spleen": 7,
+  "Accessory spleen / splenule / polysplenia": 8,
+  "Acinar infiltration areas": 9,
+  "Acinar opacities": 10,
+  "Acute rib fracture": 11,
+  "Adrenal atrophy": 12,
+  "Adrenal calcification": 13,
+  "Adrenal gland": 14,
+  "Adrenal gland calibration": 15,
+  "Adrenal gland normal": 16,
+  "Adrenal nodule": 17,
+  "Adrenal nodule (mass)": 18,
+  "Adrenal nodule / mass": 19,
+  "Adrenal thickening / hyperplasia": 20,
+  "Adrenals": 21,
+  "Air bronchograms": 22,
+  "Alveolar-interstitial density": 23,
+  "Angiomyolipoma": 24,
+  "Anterior mediastinal mass": 25,
+  "Aorta": 26,
+  "Aorta / pulmonary artery": 27,
+  "Aortic aneurysm": 28,
+  "Aortic calcification": 29,
+  "Aortic dissection / intramural hematoma": 30,
+  "Aortic stent": 31,
+  "Aortic valve calcification": 32,
+  "Aortic valve replacement": 33,
+  "Ascending aorta": 34,
+  "Ascites": 35,
+  "Atelectasis": 36,
+  "Atheroma plaques": 37,
+  "Atheroma plaques are observed in the aorta": 38,
+  "Atheromatous plaques": 39,
+  "Atherosclerotic wall calcifications": 40,
+  "Axillary lymphadenopathy": 41,
+  "Azygos fissure / lobe": 42,
+  "Azygos fissure lobe": 43,
+  "Azygos fissure variation": 44,
+  "Azygos lobe": 45,
+  "Azygos lobe / fissure": 46,
+  "Azygos lobe variation": 47,
+  "Bilateral adrenal glands": 48,
+  "Biliary drainage catheter": 49,
+  "Biliary duct dilation": 50,
+  "Biliary sludge": 51,
+  "Biliary stent / catheter / drain": 52,
+  "Bones / Spine": 53,
+  "Bones / Spine_others": 54,
+  "Bovine arch aorta": 55,
+  "Bowel wall thickening / inflammation": 56,
+  "Breast": 57,
+  "Breast & Axilla": 58,
+  "Breast implant": 59,
+  "Breast implant (intact or present)": 60,
+  "Breast mass / focal asymmetry": 61,
+  "Bronchial wall thickening": 62,
+  "Bronchiectasis": 63,
+  "Bronchopleural fistula": 64,
+  "Bronchopneumonia": 65,
+  "Bronchopneumonic infiltration": 66,
+  "Bulla": 67,
+  "Bulla / giant bulla": 68,
+  "Bullae / giant bulla": 69,
+  "CT involvement score": 70,
+  "Calcific mediastinal / hilar lymph nodes": 71,
+  "Calcified mediastinal / hilar lymph nodes": 72,
+  "Cardiac size and morphology": 73,
+  "Cardiomegaly": 74,
+  "Cardiovascular": 75,
+  "Cardiovascular_others": 76,
+  "Cavitary nodule / mass": 77,
+  "Central venous catheter": 78,
+  "Central venous catheter / PICC": 79,
+  "Centrilobular nodules / bronchiolitis pattern": 80,
+  "Cervical / supraclavicular lymphadenopathy": 81,
+  "Chest Wall": 82,
+  "Chest Wall_others": 83,
+  "Chest tube": 84,
+  "Chest tube / pleural drain": 85,
+  "Chest wall mass": 86,
+  "Chest wall soft tissue edema": 87,
+  "Chest wall soft tissue edema / hematoma": 88,
+  "Chest wall tumor invasion": 89,
+  "Cholecystectomy": 90,
+  "Cholehithiasis / gallstones": 91,
+  "Cholelithiasis": 92,
+  "Cholelithiasis / gallstones": 93,
+  "Complex renal cyst / solid renal mass": 94,
+  "Consolidation": 95,
+  "Coronary artery calcification": 96,
+  "Coronary stent or bypass graft": 97,
+  "Crazy-paving pattern": 98,
+  "Cylindrical and cystic bronchiectasis": 99,
+  "Cylindrical bronchiectasis": 100,
+  "DISH": 101,
+  "Degenerative / osseous lesions": 102,
+  "Degenerative spine changes": 103,
+  "Dextrocardia": 104,
+  "Diaphragmatic elevation": 105,
+  "Diverticulosis": 106,
+  "Effusion": 107,
+  "Emphysema": 108,
+  "Endotracheal tube": 109,
+  "Esophageal dilation": 110,
+  "Esophageal stent": 111,
+  "Esophageal wall thickening / mass": 112,
+  "Esophagus dilation": 113,
+  "Fibrotic band": 114,
+  "Focal liver lesion": 115,
+  "Focal liver lesion (nodule / mass)": 116,
+  "Focal splenic lesion": 117,
+  "Focal splenic lesion (nodule / mass)": 118,
+  "Fractures": 119,
+  "Free fluid": 120,
+  "GGO": 121,
+  "Gallbladder": 122,
+  "Gallbladder & biliary": 123,
+  "Gallbladder wall thickening": 124,
+  "Gallstones": 125,
+  "Gallstones / cholelithiasis": 126,
+  "Goiter": 127,
+  "Ground-glass opacity": 128,
+  "Ground-glass opacity (GGO)": 129,
+  "Gynecomastia": 130,
+  "Healed rib fracture": 131,
+  "Heart": 132,
+  "Heart contour": 133,
+  "Heart contour and size": 134,
+  "Heart contour and size are natural": 135,
+  "Heart contour and size are normal": 136,
+  "Heart contour and size are normal.": 137,
+  "Heart contour size": 138,
+  "Heart contour size is natural": 139,
+  "Heart contour size is natural.": 140,
+  "Heart contour size is normal": 141,
+  "Heart contour, size": 142,
+  "Heart contour, size are normal": 143,
+  "Heart contour, size are normal.": 144,
+  "Heart contour, size is normal": 145,
+  "Heart dimensions": 146,
+  "Heart dimensions and compartments": 147,
+  "Heart has a natural appearance": 148,
+  "Heart size": 149,
+  "Heart size and morphology": 150,
+  "Hemangioma": 151,
+  "Hemothorax": 152,
+  "Hepatic calcification": 153,
+  "Hepatic steatosis": 154,
+  "Hepatomegaly": 155,
+  "Hepatosteatosis": 156,
+  "Hiatal hernia": 157,
+  "Hilar lymphadenopathy": 158,
+  "Honeycomb appearance": 159,
+  "Honeycomb lung": 160,
+  "Honeycombing": 161,
+  "Horseshoe kidney": 162,
+  "Horseshoe kidney variation": 163,
+  "Hydatid cyst": 164,
+  "Hydronephrosis": 165,
+  "Hydropic gallbladder / distension": 166,
+  "IVC filter": 167,
+  "Infectious process": 168,
+  "Interlobular septal thickening": 169,
+  "Interstitial / fibrotic lung disease": 170,
+  "Interstitial lung disease-fibrosis": 171,
+  "Inverted halo sign": 172,
+  "Kidneys / urinary tract": 173,
+  "LVAD": 174,
+  "LVAD / other cardiac assist device": 175,
+  "Left atrium": 176,
+  "Linear atelectasis": 177,
+  "Liver": 178,
+  "Liver contour irregularity / cirrhosis features": 179,
+  "Liver is in normal appearance.": 180,
+  "Liver right lobe transplantation": 181,
+  "Liver transplant": 182,
+  "Lobar / segmental atelectasis": 183,
+  "Lobular kidney contours": 184,
+  "Loculated pleural effusion": 185,
+  "Lower Neck_others": 186,
+  "Lung parenchymal attenuation patterns": 187,
+  "Lungs": 188,
+  "Lungs & Airways_others": 189,
+  "Lymph nodes": 190,
+  "Lymphadenopathy": 191,
+  "Lytic bone lesion": 192,
+  "Lytic-destructive lesions": 193,
+  "Main pulmonary artery enlargement": 194,
+  "Mediastinal hematoma / fluid collection": 195,
+  "Mediastinal lymph nodes": 196,
+  "Mediastinal lymphadenopathy": 197,
+  "Mediastinal mass": 198,
+  "Mediastinal masses / cysts": 199,
+  "Mediastinal structures": 200,
+  "Mediastinum & Hila": 201,
+  "Mediastinum & Hila_others": 202,
+  "Middle / posterior mediastinal mass or cyst": 203,
+  "Mitral annular calcification": 204,
+  "Mixed osteolytic-osteosclerotic lesion": 205,
+  "Mosaic attenuation / air-trapping": 206,
+  "Motion artifact": 207,
+  "Motion artifact / suboptimal study": 208,
+  "Mucoid impaction / plugging": 209,
+  "Nasogastric / orogastric tube": 210,
+  "Nasogastric tube": 211,
+  "Neck soft tissue mass": 212,
+  "Nephrectomy": 213,
+  "Nephrectomy (kidney absent / operated)": 214,
+  "Nephrostomy catheter": 215,
+  "Neural foramina": 216,
+  "No lytic-destructive lesion": 217,
+  "No significant intrathoracic abnormality": 218,
+  "No upper abdominal free fluid-collection": 219,
+  "No upper abdominal free fluid-collection was detected in the sections.": 220,
+  "Nodular infiltrates": 221,
+  "Nodules and masses": 222,
+  "Non-acute / healed rib fracture": 223,
+  "Omental caking / peritoneal carcinomatosa": 224,
+  "Omental caking / peritoneal carcinomatosis": 225,
+  "Osteolytic bone lesion": 226,
+  "Osteopenia": 227,
+  "Osteophyte": 228,
+  "Osteophytes": 229,
+  "Osteoporosis": 230,
+  "Osteosclerotic bone lesion": 231,
+  "Others": 232,
+  "Others (devices / post-surgical / global)": 233,
+  "Others_others": 234,
+  "Pacemaker / ICD leads": 235,
+  "Pancreas": 236,
+  "Pancreatic calcification": 237,
+  "Pancreatic lipomatosis": 238,
+  "Pancreatic mass (>3 cm)": 239,
+  "Pancreatic mass / focal lesion": 240,
+  "Paraseptal emphysema": 241,
+  "Parenchymal scarring": 242,
+  "Parenchymal scarring / fibrotic band": 243,
+  "Pectus excavatum": 244,
+  "Pectus excavatum deformity": 245,
+  "Pectus excavatus": 246,
+  "Pectus excavatus anomaly": 247,
+  "Peribronchial sheath thickening": 248,
+  "Peribronchial thickening": 249,
+  "Peribronchial wall thickening": 250,
+  "Pericardial effusion": 251,
+  "Pericardial thickening": 252,
+  "Pericardial thickening / calcification": 253,
+  "Peripheral patchy ground glass densities": 254,
+  "Peritoneal carcinomatosis": 255,
+  "Pleura_others": 256,
+  "Pleural effusion": 257,
+  "Pleural nodule": 258,
+  "Pleural nodule / mass": 259,
+  "Pleural plaques": 260,
+  "Pleural thickening": 261,
+  "Pneumobilia": 262,
+  "Pneumomediastinum": 263,
+  "Pneumonia": 264,
+  "Pneumonic infiltration": 265,
+  "Pneumopericardium": 266,
+  "Pneumoperitoneum": 267,
+  "Pneumothorax": 268,
+  "Post-cholecystectomy": 269,
+  "Post-cholecystectomy (gallbladder operated / absent)": 270,
+  "Post-lobectomy / segmentectomy": 271,
+  "Post-lumpectomy / post-mastectomy change": 272,
+  "Post-mastectomy change": 273,
+  "Post-pneumonectomy": 274,
+  "Post-surgical change": 275,
+  "Post-thoracotomy change": 276,
+  "Post-thyroidectomy change": 277,
+  "Post-transplant change": 278,
+  "Postoperative spine change / hardware": 279,
+  "Postoperative stomach change": 280,
+  "Pulmonary cyst / cystic lung disease": 281,
+  "Pulmonary cysts / cystic lung disease": 282,
+  "Pulmonary embolism": 283,
+  "Pulmonary mass": 284,
+  "Pulmonary mass (>3 cm)": 285,
+  "Pulmonary nodule": 286,
+  "Pulmonary nodule (solid / PSN / GGN)": 287,
+  "Pulmonary trunk": 288,
+  "Pulmonary trunk caliber": 289,
+  "Pulmonary trunk calibration": 290,
+  "Pulmonary trunk diameter": 291,
+  "Renal artery stent": 292,
+  "Renal atrophy": 293,
+  "Renal atrophy / decreased renal size": 294,
+  "Renal calcification": 295,
+  "Renal calculi": 296,
+  "Renal calculi / nephrolithiasis": 297,
+  "Renal cyst": 298,
+  "Reticulation / intralobular thickening": 299,
+  "Schmorl nodules": 300,
+  "Schmorl's nodules": 301,
+  "Sclerotic bone lesion": 302,
+  "Scoliosis / kyphosis": 303,
+  "Septal thickening": 304,
+  "Sequela parenchymal changes": 305,
+  "Simple renal cyst": 306,
+  "Soft tissue density": 307,
+  "Soft tissue masses": 308,
+  "Spleen": 309,
+  "Spleen size": 310,
+  "Spleen sizes": 311,
+  "Splenectomy": 312,
+  "Splenomegaly": 313,
+  "Splenosis": 314,
+  "Sternal fracture": 315,
+  "Sternal hardware": 316,
+  "Study limitation / limited evaluation (non-motion)": 317,
+  "Study quality / global": 318,
+  "Subcarinal lymph nodes": 319,
+  "Subcutaneous emphysema": 320,
+  "Subsegmental / linear atelectasis": 321,
+  "Surgical hardware": 322,
+  "Surgical material": 323,
+  "Surgical suture materials": 324,
+  "Surgical sutures": 325,
+  "Suture materials": 326,
+  "Suture materials secondary to bypass surgery": 327,
+  "Syndesmophytes": 328,
+  "The left hemidiaphragm is elevated": 329,
+  "Thoracic aorta diameter": 330,
+  "Thoracic aortic aneurysm": 331,
+  "Thoracic aortic calcification": 332,
+  "Thoracic aortic diameter": 333,
+  "Thoracic aortic dilation": 334,
+  "Thoracic aortic ectasia / dilation": 335,
+  "Thoracic aortic ectasia / dilation (non-aneurysmal)": 336,
+  "Thoracic esophageal calibration": 337,
+  "Thoracic esophageal dilation": 338,
+  "Thoracic esophageal wall thickening": 339,
+  "Thoracic esophageal wall thickening / mass": 340,
+  "Thoracic esophagus": 341,
+  "Thoracic esophagus calibration": 342,
+  "Thoracic kyphosis": 343,
+  "Thoracic vertebral corpus heights": 344,
+  "Thoracic vertebral corpus heights, alignments and densities are normal": 345,
+  "Thoracic vertebral corpus heights, alignments and densities are normal.": 346,
+  "Thymic remnant / hyperplasia": 347,
+  "Thyroid enlargement (goiter)": 348,
+  "Thyroid nodule": 349,
+  "Trachea": 350,
+  "Trachea & Airways": 351,
+  "Trachea and main bronchi are open": 352,
+  "Tracheal / bronchial wall thickening": 353,
+  "Tracheal diverticulum": 354,
+  "Tracheal stenosis": 355,
+  "Tracheal stenosis / malacia": 356,
+  "Tracheal wall thickening": 357,
+  "Tracheobronchopathy osteochondroplastica": 358,
+  "Tracheomegaly": 359,
+  "Tracheostomy tube": 360,
+  "Traction bronchiectasis": 361,
+  "Traction bronchiectasis / bronchiolectasis": 362,
+  "Tree-in-bud": 363,
+  "Tubes, catheters, and support devices": 364,
+  "Upper Abdomen": 365,
+  "Upper Abdomen_others": 366,
+  "Upper abdominal free fluid-collection": 367,
+  "Upper abdominal vessels": 368,
+  "Valves and cardiac devices": 369,
+  "Vertebral compression fracture": 370,
+  "Vertebral corpus height": 371,
+  "Vertebral corpus heights": 372,
+  "Vertebral corpus heights are preserved": 373,
+  "Vertebral corpus heights are preserved.": 374,
+  "Vertebral corpus heights preserved": 375,
+  "Vertebral corpus heights, alignments and densities": 376,
+  "Vertebral corpus heights, alignments and densities within the sections are normal": 377,
+  "Vertebral fracture": 378,
+  "Vertebral hemangioma": 379,
+  "Viral pneumonia": 380,
+  "Volume loss / hyperinflation": 381,
+  "azygos fissure variation": 382,
+  "tree-in-bud": 383,
+  "<UNK>": 384
+}