"""HEMATO v1.0 builder. Reads raw sources (raw/), writes dataset//// and manifest.csv. Configs: cls (bone marrow + peripheral blood cell types), all (C-NMC ALL vs normal), malaria (NIH), det (BCCD, YOLO). Images are copied byte-for-byte when already JPEG/PNG; BMP (C-NMC) is converted losslessly to PNG.""" import csv, glob, io, os, random, re, shutil, zipfile from collections import defaultdict import pyarrow.parquet as pq from PIL import Image RAW, OUT = 'raw', 'dataset' SEED = 42 RATIOS = (('train', .70), ('val', .15), ('test', .15)) # ---- unified cell-type vocabulary (cls) ---------------------------------------------------------- # folder name: (Turkish name, bone-marrow code (Matek 2021), PBC filename prefix (Acevedo 2020)) CLASSES = { 'Blast': ('Blast', 'BLA', None), 'Promyelocyte': ('Promiyelosit', 'PMO', 'PMY'), 'Myelocyte': ('Miyelosit', 'MYB', 'MY'), 'Metamyelocyte': ('Metamiyelosit', 'MMZ', 'MMY'), 'Band_neutrophil': ('Çomak nötrofil', 'NGB', 'BNE'), 'Segmented_neutrophil': ('Segmentli nötrofil', 'NGS', 'SNE'), 'Eosinophil': ('Eozinofil', 'EOS', 'EO'), 'Basophil': ('Bazofil', 'BAS', 'BA'), 'Lymphocyte': ('Lenfosit', 'LYT', 'LY'), 'Monocyte': ('Monosit', 'MON', 'MO'), 'Proerythroblast': ('Proeritroblast', 'PEB', None), 'Erythroblast': ('Eritroblast', 'EBO', 'ERB'), 'Plasma_cell': ('Plazma hücresi', 'PLM', None), 'Hairy_cell': ('Tüylü hücre (hairy cell)', 'HAC', None), 'Platelet': ('Trombosit', None, 'PLATELET'), 'Artefact': ('Artefakt', 'ART', None), 'Not_identifiable': ('Tanımlanamayan hücre', 'NIF', None), } BM_NAMES = ['ABE', 'ART', 'BAS', 'BLA', 'EBO', 'EOS', 'FGC', 'HAC', 'KSC', 'LYI', 'LYT', 'MMZ', 'MON', 'MYB', 'NGB', 'NGS', 'NIF', 'OTH', 'PEB', 'PLM', 'PMO'] BM2CLS = {v[1]: k for k, v in CLASSES.items() if v[1]} PBC2CLS = {v[2]: k for k, v in CLASSES.items() if v[2]} rows, excluded = [], [] def add(config, split, cls, fname, data, **meta): d = os.path.join(OUT, config, split, cls); os.makedirs(d, exist_ok=True) with open(os.path.join(d, fname), 'wb') as f: f.write(data) rows.append(dict(config=config, split=split, label=cls, file=f'{config}/{split}/{cls}/{fname}', **meta)) def split_items(items, key=lambda x: x): """stratified random split of a list (per class) -> dict split->list""" rnd = random.Random(SEED); items = sorted(items, key=key); rnd.shuffle(items) n = len(items); a = round(n * .70); b = a + round(n * .15) return {'train': items[:a], 'val': items[a:b], 'test': items[b:]} def split_groups(groups_by_label): """patient-level split: groups (patients) are shuffled per label and assigned 70/15/15""" out = {} for lab, groups in groups_by_label.items(): for s, gs in split_items(sorted(groups)).items(): for g in gs: out[g] = s return out # ---- 1. bone marrow (Matek et al. 2021, MLL) ----------------------------------------------------- BM_FILES = sorted(glob.glob(f'{RAW}/bm/data/*.parquet')) bm = defaultdict(list) # pass 1: labels + paths only (no image bytes -> low memory) for f in BM_FILES: pf = pq.ParquetFile(f) for b in pf.iter_batches(batch_size=2000, columns=['image', 'label']): d = b.to_pydict() for im, lab in zip(d['image'], d['label']): code = BM_NAMES[lab] if code not in BM2CLS: excluded.append(('bone_marrow', im['path'], code, 'class too small / heterogeneous')); continue bm[BM2CLS[code]].append(im['path']) bm_split = {} for cls, paths in bm.items(): for split, ps in split_items(paths).items(): for p in ps: bm_split[p] = (split, cls) for f in BM_FILES: # pass 2: stream image bytes batch by batch for b in pq.ParquetFile(f).iter_batches(batch_size=500, columns=['image', 'label']): d = b.to_pydict() for im, lab in zip(d['image'], d['label']): if im['path'] not in bm_split: continue split, cls = bm_split[im['path']] add('cls', split, cls, f'BM_{im["path"]}', im['bytes'], source='bone_marrow_mll', original_label=BM_NAMES[lab], original_file=im['path'], patient_id='') print('bone marrow done', len(bm_split)) # ---- 2. peripheral blood (PBC, Acevedo et al. 2020) ---------------------------------------------- z = zipfile.ZipFile(f'{RAW}/pbc/PBC_dataset_normal_DIB.zip') pbc = defaultdict(list) for n in z.namelist(): if n.startswith('__MACOSX') or not n.lower().endswith('.jpg'): continue fn = n.split('/')[-1]; pre = fn.split('_')[0] if pre not in PBC2CLS: excluded.append(('pbc', n, pre, 'subtype not specified (IG / NEUTROPHIL)')); continue pbc[PBC2CLS[pre]].append((n, fn, pre)) for cls, items in pbc.items(): for split, its in split_items(items, key=lambda x: x[0]).items(): for n, fn, pre in its: add('cls', split, cls, f'PBC_{fn}', z.read(n), source='pbc_acevedo', original_label=pre, original_file=n, patient_id='') print('pbc done', sum(len(v) for v in pbc.values())) # ---- 3. C-NMC 2019 ALL vs normal (patient-level split) ------------------------------------------- CN_FILES = sorted(glob.glob(f'{RAW}/cnmc/data/*.parquet')) g = defaultdict(set) t = pq.read_table(CN_FILES, columns=['subject_id', 'label']).to_pydict() for sid, lab in zip(t['subject_id'], t['label']): g['ALL' if lab == 'cancer' else 'Normal'].add(sid) sp = split_groups(g) n_cn = 0 for f in CN_FILES: for b in pq.ParquetFile(f).iter_batches(batch_size=300, columns=['subject_id', 'label', 'image', 'original_image_name']): d = b.to_pydict() for sid, lab, im, name in zip(d['subject_id'], d['label'], d['image'], d['original_image_name']): lab = 'ALL' if lab == 'cancer' else 'Normal' try: buf = io.BytesIO(); Image.open(io.BytesIO(im['bytes'])).save(buf, 'PNG', optimize=True) except Exception as e: excluded.append(('c_nmc_2019', name, lab, f'corrupt image ({type(e).__name__})')); continue add('all', sp[sid], lab, f'CNMC_{os.path.splitext(name)[0]}.png', buf.getvalue(), source='c_nmc_2019', original_label=lab, original_file=name, patient_id=f'CNMC_{sid}') n_cn += 1 print('cnmc done', n_cn) # ---- 4. NIH malaria (patient-level split) -------------------------------------------------------- z = zipfile.ZipFile(f'{RAW}/malaria/cell_images.zip') mal = [] for n in z.namelist(): if not n.endswith('.png'): continue lab = 'Parasitized' if '/Parasitized/' in n else 'Uninfected' fn = n.split('/')[-1]; pid = re.split(r'[Tt]hin', fn)[0].rstrip('_') mal.append((pid, lab, n, fn)) pat_lab = defaultdict(set) for pid, lab, _, _ in mal: pat_lab[pid].add(lab) g = defaultdict(set) for pid, labs in pat_lab.items(): g['infected_patient' if 'Parasitized' in labs else 'healthy_patient'].add(pid) sp = split_groups(g) for pid, lab, n, fn in mal: add('malaria', sp[pid], lab, f'NIH_{fn}', z.read(n), source='nih_malaria', original_label=lab, original_file=n, patient_id=f'NIH_{pid}') print('malaria done', len(mal)) # ---- 5. BCCD detection (YOLO, original split kept) ----------------------------------------------- for s_in, s_out in (('train', 'train'), ('valid', 'val'), ('test', 'test')): for img in sorted(glob.glob(f'{RAW}/bccd/{s_in}/images/*')): base = os.path.splitext(os.path.basename(img))[0] lbl = f'{RAW}/bccd/{s_in}/labels/{base}.txt' for sub, src in (('images', img), ('labels', lbl)): d = os.path.join(OUT, 'det', sub, s_out); os.makedirs(d, exist_ok=True) shutil.copy(src, d) rows.append(dict(config='det', split=s_out, label='', file=f'det/images/{s_out}/{os.path.basename(img)}', source='bccd', original_label='', original_file=img.split('/', 2)[-1], patient_id='')) with open(os.path.join(OUT, 'det', 'data.yaml'), 'w') as f: f.write('path: .\ntrain: images/train\nval: images/val\ntest: images/test\nnc: 3\nnames: [Platelets, RBC, WBC]\n') print('bccd done') with open('manifest.csv', 'w', newline='') as f: w = csv.DictWriter(f, fieldnames=list(rows[0])); w.writeheader(); w.writerows(rows) with open('excluded.csv', 'w', newline='') as f: w = csv.writer(f); w.writerow(['source', 'file', 'original_label', 'reason']); w.writerows(excluded) print('rows', len(rows), 'excluded', len(excluded))