File size: 13,475 Bytes
4ab0193
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
import csv
import json
import os
import random
from collections import defaultdict
from multiprocessing import Pool

from PIL import Image

BASE_DIR    = "./Chest_imagenome"
DATA_DIR    = os.path.join(BASE_DIR, "gold")
BBOX_CSV    = os.path.join(DATA_DIR, "gold_bbox_coordinate_annotations_1000images.csv")
SENTENCES   = os.path.join(DATA_DIR, "gold_all_sentences_500pts_1000studies.txt")
SILVER_DIR  = os.path.join(BASE_DIR, "silver")
SG_DIR      = os.path.join(SILVER_DIR, "scene_graph")
SPLIT_DIR   = os.path.join(SILVER_DIR, "split")
MIMIC_DIR   = "./mimic_original/2.0.0/files"

GOLD_CROP_DIR   = os.path.join(BASE_DIR, "gold_crop")
SILVER_CROP_DIR = os.path.join(BASE_DIR, "silver_sample_crop")

OUTPUT_TRAIN_JSON = os.path.join(BASE_DIR, "Chest_imagenome_train.json")
OUTPUT_TEST_JSON  = os.path.join(BASE_DIR, "Chest_imagenome_test.json")

TARGET_TOTAL     = 1_000_000
MIN_SIZE         = 28
NUM_WORKERS_GOLD   = 8
NUM_WORKERS_SILVER = 16
MAX_TEST_SAMPLES = 1000
RANDOM_SEED      = 42
SEED             = 42

TRAIN_IMAGE_ROOT = MIMIC_DIR
TRAIN_CROP_ROOT  = BASE_DIR
TEST_IMAGE_ROOT  = MIMIC_DIR
TEST_CROP_ROOT   = BASE_DIR


def crop_with_min_size(img, x_min, y_min, x_max, y_max, min_size=MIN_SIZE):
    cx = (x_min + x_max) // 2
    cy = (y_min + y_max) // 2
    w  = max(x_max - x_min, min_size)
    h  = max(y_max - y_min, min_size)
    x_min = max(0, cx - w // 2); x_max = min(img.width,  x_min + w); x_min = max(0, x_max - w)
    y_min = max(0, cy - h // 2); y_max = min(img.height, y_min + h); y_min = max(0, y_max - h)
    return img.crop((x_min, y_min, x_max, y_max))

def normalize_path(path):
    return str(path).replace("\\", "/")

def remove_root_prefix(path, root):
    path = normalize_path(path)
    root = normalize_path(root).rstrip("/")
    if path.startswith(root + "/"):
        return path[len(root) + 1:]
    elif path == root:
        return ""
    return path


def process_gold_image(args):
    image_id, regions, class_to_id, img_path = args

    try:
        img = Image.open(img_path)
    except Exception as e:
        return None, f"open {image_id}: {e}"

    map_rows = []
    for bbox_name, box in regions:
        class_id  = class_to_id[bbox_name]
        crop_path = os.path.join(GOLD_CROP_DIR, f"{image_id}_{class_id}_0.png")
        try:
            crop_with_min_size(img, *box).save(crop_path)
        except Exception as e:
            return None, f"crop {image_id} {bbox_name}: {e}"

        map_rows.append({
            "image_path":       img_path,
            "class_name":       bbox_name,
            "crop_image_paths": [crop_path],
        })

    return map_rows, None


def run_gold():
    os.makedirs(GOLD_CROP_DIR, exist_ok=True)

    img_to_path = {}
    with open(SENTENCES, newline="") as f:
        for row in csv.DictReader(f, delimiter="\t"):
            image_id = row["image_id"]
            if image_id in img_to_path:
                continue
            patient_id = row["patient_id"]
            study_id   = row["study_id"]
            jpg_name   = image_id.replace(".dcm", ".jpg")
            img_to_path[image_id] = os.path.join(
                MIMIC_DIR, f"p{patient_id[:2]}", f"p{patient_id}", f"s{study_id}", jpg_name
            )
    print(f"[gold] image path mapping: {len(img_to_path)}")

    image_regions = defaultdict(list)
    all_names = set()
    with open(BBOX_CSV, newline="") as f:
        for row in csv.DictReader(f):
            image_id  = row["image_id"]
            bbox_name = row["bbox_name"]
            box = (
                max(0, int(float(row["original_x1"]))),
                max(0, int(float(row["original_y1"]))),
                max(0, int(float(row["original_x2"]))),
                max(0, int(float(row["original_y2"]))),
            )
            image_regions[image_id].append((bbox_name, box))
            all_names.add(bbox_name)

    class_to_id = {name: idx for idx, name in enumerate(sorted(all_names))}
    print(f"[gold] regions: {len(class_to_id)}  images: {len(image_regions)}")

    tasks = []
    for image_id, regions in image_regions.items():
        if image_id not in img_to_path:
            continue
        img_id_short = image_id.replace(".dcm", "")
        tasks.append((img_id_short, regions, class_to_id, img_to_path[image_id]))

    print(f"[gold] tasks: {len(tasks)}  workers: {NUM_WORKERS_GOLD}")

    records = []
    errors  = 0
    with Pool(NUM_WORKERS_GOLD) as pool:
        for i, (map_rows, err) in enumerate(
            pool.imap_unordered(process_gold_image, tasks, chunksize=4), start=1
        ):
            if err:
                print(f"  ERROR: {err}"); errors += 1
            elif map_rows:
                records.extend(map_rows)
            if i % 200 == 0:
                print(f"  [gold {i}/{len(tasks)}] errors={errors}")

    print(f"[gold] done. records={len(records)} errors={errors}")
    return records


def scan_one_json(args):
    sg_path, dicom_id = args
    try:
        with open(sg_path) as f:
            sg = json.load(f)
    except Exception:
        return None
    names = [obj["bbox_name"] for obj in sg.get("objects", []) if obj.get("original_x1") is not None]
    return (dicom_id, names) if names else None


def process_silver_image(args):
    dicom_id, img_path, sg_path, split, selected_names, class_to_id = args

    try:
        with open(sg_path) as f:
            sg = json.load(f)
    except Exception as e:
        return None, f"json {dicom_id}: {e}"

    obj_map = {obj["bbox_name"]: obj for obj in sg.get("objects", [])}

    try:
        img = Image.open(img_path)
    except Exception as e:
        return None, f"open {dicom_id}: {e}"

    map_rows = []
    for bbox_name in selected_names:
        obj = obj_map.get(bbox_name)
        if obj is None:
            continue
        ox1, oy1, ox2, oy2 = obj.get("original_x1"), obj.get("original_y1"), obj.get("original_x2"), obj.get("original_y2")
        if any(v is None for v in (ox1, oy1, ox2, oy2)):
            continue

        box      = (max(0, int(ox1)), max(0, int(oy1)), max(0, int(ox2)), max(0, int(oy2)))
        class_id = class_to_id[bbox_name]
        crop_path = os.path.join(SILVER_CROP_DIR, f"{dicom_id}_{class_id}_0.png")

        try:
            crop_with_min_size(img, *box).save(crop_path)
        except Exception as e:
            return None, f"crop {dicom_id} {bbox_name}: {e}"

        map_rows.append({
            "image_path":       img_path,
            "class_name":       bbox_name,
            "split":            split,
            "crop_image_paths": [crop_path],
        })

    return map_rows, None


def run_silver():
    os.makedirs(SILVER_CROP_DIR, exist_ok=True)
    random.seed(SEED)

    id_map = {}
    for split_name, csv_name in [("training", "train.csv"), ("validation", "valid.csv"), ("test", "test.csv")]:
        with open(os.path.join(SPLIT_DIR, csv_name), newline="") as f:
            for row in csv.DictReader(f):
                id_map[row["dicom_id"]] = (row["subject_id"], row["study_id"], split_name)
    print(f"[silver] split mapping: {len(id_map)}")

    sg_files  = [fn for fn in os.listdir(SG_DIR) if fn.endswith("_SceneGraph.json")]
    scan_args = [
        (os.path.join(SG_DIR, fn), fn.replace("_SceneGraph.json", ""))
        for fn in sg_files
        if fn.replace("_SceneGraph.json", "") in id_map
    ]
    print(f"[silver] scanning {len(scan_args)} JSONs...")

    anatomy_to_images = defaultdict(list)
    with Pool(NUM_WORKERS_SILVER) as pool:
        for i, result in enumerate(pool.imap_unordered(scan_one_json, scan_args, chunksize=256)):
            if result:
                dicom_id, names = result
                for name in names:
                    anatomy_to_images[name].append(dicom_id)
            if (i + 1) % 50000 == 0:
                print(f"  scanned {i+1}/{len(scan_args)}")

    n_anatomy  = len(anatomy_to_images)
    per_anatomy = TARGET_TOTAL // n_anatomy
    print(f"[silver] anatomy types: {n_anatomy}  per anatomy: {per_anatomy}")

    image_selected = defaultdict(set)
    for bbox_name, img_list in sorted(anatomy_to_images.items()):
        n_sample = min(per_anatomy, len(img_list))
        for dicom_id in random.sample(img_list, n_sample):
            image_selected[dicom_id].add(bbox_name)

    class_to_id = {name: idx for idx, name in enumerate(sorted(anatomy_to_images.keys()))}

    tasks = []
    for dicom_id, selected_names in image_selected.items():
        subject_id, study_id, split = id_map[dicom_id]
        img_path = os.path.join(MIMIC_DIR, f"p{subject_id[:2]}", f"p{subject_id}", f"s{study_id}", f"{dicom_id}.jpg")
        sg_path  = os.path.join(SG_DIR, f"{dicom_id}_SceneGraph.json")
        tasks.append((dicom_id, img_path, sg_path, split, list(selected_names), class_to_id))

    print(f"[silver] tasks: {len(tasks)}  workers: {NUM_WORKERS_SILVER}")

    records = []
    errors  = 0
    with Pool(NUM_WORKERS_SILVER) as pool:
        for i, (map_rows, err) in enumerate(
            pool.imap_unordered(process_silver_image, tasks, chunksize=32), start=1
        ):
            if err:
                print(f"  ERROR: {err}"); errors += 1
            elif map_rows:
                records.extend(map_rows)
            if i % 10000 == 0:
                print(f"  [silver {i}/{len(tasks)}] errors={errors}")

    print(f"[silver] done. records={len(records)} errors={errors}")
    return records


def convert_record(record, image_root, crop_root):
    img_path_rel   = remove_root_prefix(record["image_path"], image_root)
    crop_paths_rel = [remove_root_prefix(p, crop_root) for p in record.get("crop_image_paths", [])]
    tgt_img_path   = crop_paths_rel if len(crop_paths_rel) <= 1 else [crop_paths_rel]

    return {
        "qry_inst":     "<|image_1|> Locate the specific region that corresponds to the provided text description.",
        "qry_text":     record["class_name"],
        "qry_img_path": img_path_rel,
        "tgt_inst":     "Match the target",
        "tgt_text":     ["<|image_1|>\n"],
        "tgt_img_path": tgt_img_path,
    }


def deduplicate(data):
    seen   = set()
    result = []
    for s in data:
        key = (s.get("qry_img_path", ""), s.get("qry_text", ""))
        if key not in seen:
            seen.add(key)
            result.append(s)
    removed = len(data) - len(result)
    return result, removed


def sample_test_data(data, max_samples, seed=RANDOM_SEED):
    if len(data) <= max_samples:
        return data

    rng = random.Random(seed)

    class_to_samples = defaultdict(list)
    for s in data:
        class_to_samples[str(s.get("qry_text", "")).strip()].append(s)
    for v in class_to_samples.values():
        rng.shuffle(v)

    class_names    = list(class_to_samples.keys())
    sampled        = []
    used_paths     = set()
    used_ids       = set()

    still_has = True
    while still_has and len(sampled) < max_samples:
        still_has = False
        rng.shuffle(class_names)
        for cls in class_names:
            candidates = class_to_samples[cls]
            if not candidates:
                continue
            still_has = True
            chosen_idx = next(
                (i for i, s in enumerate(candidates)
                 if s.get("qry_img_path") not in used_paths and id(s) not in used_ids),
                None
            )
            if chosen_idx is None:
                chosen_idx = next(
                    (i for i, s in enumerate(candidates) if id(s) not in used_ids), None
                )
            if chosen_idx is None:
                continue
            chosen = candidates.pop(chosen_idx)
            sampled.append(chosen)
            used_ids.add(id(chosen))
            used_paths.add(chosen.get("qry_img_path", ""))
            if len(sampled) >= max_samples:
                break

    if len(sampled) < max_samples:
        remaining = [s for v in class_to_samples.values() for s in v if id(s) not in used_ids]
        unique_r  = [s for s in remaining if s.get("qry_img_path") not in used_paths]
        dup_r     = [s for s in remaining if s.get("qry_img_path") in used_paths]
        rng.shuffle(unique_r); rng.shuffle(dup_r)
        for s in unique_r + dup_r:
            if len(sampled) >= max_samples:
                break
            sampled.append(s)

    rng.shuffle(sampled)
    return sampled[:max_samples]


def main():
    gold_records   = run_gold()
    silver_records = run_silver()

    test_data  = [convert_record(r, TEST_IMAGE_ROOT,  TEST_CROP_ROOT)  for r in gold_records]
    train_data = [convert_record(r, TRAIN_IMAGE_ROOT, TRAIN_CROP_ROOT) for r in silver_records]

    rng = random.Random(RANDOM_SEED)
    rng.shuffle(train_data)
    rng.shuffle(test_data)

    train_data, rm_train = deduplicate(train_data)
    test_data,  rm_test  = deduplicate(test_data)
    print(f"train dedup removed: {rm_train}  remaining: {len(train_data)}")
    print(f"test  dedup removed: {rm_test}   remaining: {len(test_data)}")

    original_test_count = len(test_data)
    test_data = sample_test_data(test_data, MAX_TEST_SAMPLES, RANDOM_SEED)
    print(f"test samples: {original_test_count} -> {len(test_data)}")

    for path, data in [(OUTPUT_TRAIN_JSON, train_data), (OUTPUT_TEST_JSON, test_data)]:
        os.makedirs(os.path.dirname(path), exist_ok=True)
        with open(path, "w", encoding="utf-8") as f:
            json.dump(data, f, ensure_ascii=False, indent=2)
        print(f"Saved: {path}")


if __name__ == "__main__":
    main()