# %% [code] """ RSNA Knee Abnormality Detection: v30 INFERENCE Loads best_model_v29.pth (val AUC 0.742) and runs inference on ALL test studies. Outputs submission.csv to /kaggle/working. """ import os import sys import glob import numpy as np import pandas as pd import torch import torch.nn as nn from torch.utils.data import Dataset, DataLoader from pydicom.pixel_data_handlers.util import apply_voi_lut import pydicom import cv2 print(f"PyTorch: {torch.__version__}, CUDA: {torch.cuda.is_available()}") device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') BASE = '/kaggle/input/competitions/rsna-knee-abnormality-detection' TARGETS = ['ACL', 'MCL', 'Medial Meniscus', 'Lateral Meniscus', 'Medial OA', 'Lateral OA', 'PF OA', 'Effusion', 'Synovitis', "Baker's", 'Contusion', 'Fracture'] class CFG: img_size = 128 num_slices = 3 num_planes = 2 def find_csv(name, root='/kaggle/input'): for dirpath, _, files in os.walk(root): if name in files: return os.path.join(dirpath, name) raise FileNotFoundError(name) WEIGHTS = find_csv('best_model_v29.pth') print('[Model] weights:', WEIGHTS) # ------------- Model (same as training) ------------- class KneeModel(nn.Module): def __init__(self): super().__init__() from torchvision.models import resnet18 base = resnet18(weights=None) fp = find_csv('resnet18-f37072fd.pth') base.load_state_dict(torch.load(fp, map_location='cpu', weights_only=True)) base.fc = nn.Identity() self.backbone = base self.att = nn.Sequential(nn.Linear(512, 64), nn.ReLU()) self.gate = nn.Linear(64, 1) self.head = nn.Sequential( nn.Linear(512, 256), nn.ReLU(), nn.Dropout(0.2), nn.Linear(256, 12)) def forward(self, x): x = x.float() B = x.size(0) x = x.view(B, -1, 3, x.size(-2), x.size(-1)) K = x.size(1) s = x.view(B * K, 3, x.size(-2), x.size(-1)) feats = self.backbone(s).view(B, K, -1) a = self.att(feats) g = torch.softmax(self.gate(a), dim=1) fused = (feats * g).sum(dim=1) return self.head(fused) # ------------- Image loading ------------- def prep_stack(files, max_slices=CFG.num_slices, img_size=CFG.img_size): imgs = [] for f in files: try: dcm = pydicom.dcmread(f, stop_before_pixels=False, force=True) px = dcm.pixel_array.astype(np.float32) px = apply_voi_lut(px, dcm) if px.max() > px.min(): px = (px - px.min()) / (px.max() - px.min()) else: continue px = np.clip(px, 0, 1) px = cv2.resize(px, (img_size, img_size), interpolation=cv2.INTER_AREA) imgs.append(px) except Exception: continue if not imgs: return None n = len(imgs) idx = np.linspace(0, n - 1, max_slices, dtype=int) if n >= max_slices else np.arange(n) stack = np.stack([imgs[i] for i in idx], axis=0) rgb = np.repeat(stack[:, None, :, :], 3, axis=1) return torch.from_numpy(rgb.astype(np.float32)) # ------------- Dataset ------------- class KneeDataset(Dataset): def __init__(self, df, series_df, img_dir): self.ids = df['StudyInstanceUID'].astype(str).str.strip().values self.img_dir = img_dir self.series_map = {} self.fluid = {} for _, r in series_df.iterrows(): s_id = str(r.iloc[0]).strip() ser = str(r.iloc[1]).strip() self.series_map.setdefault(s_id, []) if ser not in self.series_map[s_id]: self.series_map[s_id].append(ser) self.fluid[ser] = int(r.iloc[2]) if 'Fluid_Sensitive' in series_df.columns else 0 print(f'[Dataset] {len(self.ids)} studies, {len(self.series_map)} mapped') def __len__(self): return len(self.ids) def __getitem__(self, i): sid = str(self.ids[i]).strip() series_list = self.series_map.get(sid, []) fluid = [s for s in series_list if self.fluid.get(s, 0) == 1] non_fluid = [s for s in series_list if self.fluid.get(s, 0) == 0] selected = [] for pool in [fluid, non_fluid]: for s in pool: if len(selected) >= CFG.num_planes: break selected.append(s) if len(selected) >= CFG.num_planes: break if not selected: selected = series_list[:CFG.num_planes] stacks = [] zero = torch.zeros((CFG.num_slices, 3, CFG.img_size, CFG.img_size)) for s in selected: path = os.path.join(self.img_dir, sid, s) if not os.path.exists(path): stacks.append(zero) continue files = sorted(glob.glob(os.path.join(path, '*.dcm'))) if not files: stacks.append(zero) continue t = prep_stack(files) stacks.append(t if t is not None else zero) while len(stacks) < CFG.num_planes: stacks.append(zero) x = torch.stack(stacks[:CFG.num_planes], dim=0) return x.view(-1, 3, CFG.img_size, CFG.img_size) # ------------- Main ------------- def main(): test_df = pd.read_csv(os.path.join(BASE, 'test.csv')) print(f'[Data] test studies: {len(test_df)}') img_dir = os.path.join(BASE, 'train_series') test_series = pd.read_csv(find_csv('test_series.csv')) print(f'[Data] test series rows: {len(test_series)}') ds = KneeDataset(test_df, test_series, img_dir) loader = DataLoader(ds, batch_size=16, shuffle=False, num_workers=0) model = KneeModel().to(device) model.load_state_dict(torch.load(WEIGHTS, map_location=device)) model.eval() print('[Inference] running...') preds = [] with torch.no_grad(): for x in loader: x = x.to(device) preds.append(torch.sigmoid(model(x)).cpu().numpy()) preds = np.vstack(preds) print(f'[Inference] shape={preds.shape}, mean={preds.mean():.4f}') sub = pd.DataFrame(preds, columns=TARGETS) sub.insert(0, 'StudyInstanceUID', test_df['StudyInstanceUID'].astype(str).str.strip().values[:len(sub)]) sub.to_csv('/kaggle/working/submission.csv', index=False) print('[Saved] /kaggle/working/submission.csv') print(sub.to_string()) sys.stdout.flush() if __name__ == '__main__': main()