|
|
import argparse |
|
|
import os |
|
|
import sys |
|
|
from types import SimpleNamespace |
|
|
from typing import Any |
|
|
|
|
|
import yaml |
|
|
import numpy as np |
|
|
import torch |
|
|
import nibabel as nib |
|
|
from torch.utils.data import DataLoader |
|
|
|
|
|
|
|
|
os.environ.setdefault("MONAI_SKIP_SUBMODULES", "1") |
|
|
from monai.inferers import sliding_window_inference |
|
|
|
|
|
sys.path.append(os.path.join(os.path.dirname(__file__), "src")) |
|
|
|
|
|
from gliomasam3_moe.models.gliomasam3_moe import GliomaSAM3_MoE |
|
|
from gliomasam3_moe.data.brats_dataset import BraTSDataset, SegMambaNPZDataset |
|
|
from gliomasam3_moe.data.transforms_segmamba_like import get_infer_transforms |
|
|
from gliomasam3_moe.utils.brats_regions import regions_to_label |
|
|
from gliomasam3_moe.utils.postprocess import remove_small_components |
|
|
|
|
|
|
|
|
def _to_namespace(obj: Any): |
|
|
if isinstance(obj, dict): |
|
|
return SimpleNamespace(**{k: _to_namespace(v) for k, v in obj.items()}) |
|
|
return obj |
|
|
|
|
|
|
|
|
def load_config(path: str) -> SimpleNamespace: |
|
|
with open(path, "r") as f: |
|
|
cfg = yaml.safe_load(f) |
|
|
return _to_namespace(cfg) |
|
|
|
|
|
|
|
|
def _get_affine(meta_dict): |
|
|
if meta_dict is None: |
|
|
return np.eye(4) |
|
|
affine = meta_dict.get("affine", None) |
|
|
if isinstance(affine, torch.Tensor): |
|
|
affine = affine.detach().cpu().numpy() |
|
|
if isinstance(affine, np.ndarray) and affine.ndim == 3: |
|
|
affine = affine[0] |
|
|
if affine is None: |
|
|
affine = np.eye(4) |
|
|
return affine |
|
|
|
|
|
|
|
|
def save_nifti(path: str, arr: np.ndarray, affine: np.ndarray): |
|
|
img = nib.Nifti1Image(arr, affine) |
|
|
nib.save(img, path) |
|
|
|
|
|
|
|
|
def save_segmamba_3c(path: str, arr_3c: np.ndarray, affine: np.ndarray | None = None): |
|
|
"""Save 3-channel mask for SegMamba metrics. |
|
|
|
|
|
Expected input: [3, D, H, W], saved as 4D NIfTI (D,H,W,3). |
|
|
""" |
|
|
if affine is None: |
|
|
affine = np.eye(4) |
|
|
if arr_3c.ndim != 4 or arr_3c.shape[0] != 3: |
|
|
raise ValueError(f"expected (3,D,H,W), got {arr_3c.shape}") |
|
|
arr = arr_3c.transpose(1, 2, 3, 0) |
|
|
save_nifti(path, arr.astype(np.uint8), affine) |
|
|
|
|
|
|
|
|
def main(): |
|
|
parser = argparse.ArgumentParser() |
|
|
parser.add_argument("--config", type=str, default="configs/train.yaml") |
|
|
parser.add_argument("--input", type=str, required=True, help="Case folder or root folder.") |
|
|
parser.add_argument("--output", type=str, default="./prediction_results/segmamba") |
|
|
parser.add_argument("--checkpoint", type=str, required=True) |
|
|
args = parser.parse_args() |
|
|
|
|
|
cfg = load_config(args.config) |
|
|
device = torch.device(cfg.device if torch.cuda.is_available() else "cpu") |
|
|
|
|
|
model = GliomaSAM3_MoE(**cfg.model.__dict__).to(device) |
|
|
ckpt = torch.load(args.checkpoint, map_location="cpu") |
|
|
model.load_state_dict(ckpt["model"], strict=True) |
|
|
model.eval() |
|
|
data_format = getattr(cfg.data, "format", "nifti") |
|
|
|
|
|
input_path = args.input |
|
|
if data_format == "segmamba_npz": |
|
|
if not os.path.isdir(input_path): |
|
|
raise ValueError("Input must be a directory containing *.npz files.") |
|
|
ensure_npy = bool(getattr(cfg.data, "segmamba_unpack", True)) |
|
|
dataset = SegMambaNPZDataset( |
|
|
data_dir=input_path, |
|
|
test=True, |
|
|
ensure_npy=ensure_npy, |
|
|
map_et_to_4=True, |
|
|
) |
|
|
loader = DataLoader(dataset, batch_size=1, shuffle=False, num_workers=0) |
|
|
else: |
|
|
if os.path.isdir(input_path): |
|
|
has_modalities = any( |
|
|
os.path.isfile(os.path.join(input_path, m + ".nii.gz")) or os.path.isfile(os.path.join(input_path, m + ".nii")) |
|
|
for m in cfg.data.modalities |
|
|
) |
|
|
if has_modalities: |
|
|
root_dir = os.path.dirname(input_path) |
|
|
case_ids = [os.path.basename(input_path)] |
|
|
else: |
|
|
root_dir = input_path |
|
|
case_ids = None |
|
|
else: |
|
|
raise ValueError("Input must be a directory.") |
|
|
|
|
|
image_keys = [f"image{i}" for i in range(len(cfg.data.modalities))] |
|
|
transforms = get_infer_transforms(cfg, image_keys=image_keys) |
|
|
dataset = BraTSDataset( |
|
|
root_dir=root_dir, |
|
|
modalities=cfg.data.modalities, |
|
|
seg_name=cfg.data.seg_name, |
|
|
transforms=transforms, |
|
|
include_label=False, |
|
|
case_ids=case_ids, |
|
|
image_keys=image_keys, |
|
|
) |
|
|
loader = DataLoader(dataset, batch_size=1, shuffle=False, num_workers=0) |
|
|
|
|
|
os.makedirs(args.output, exist_ok=True) |
|
|
with torch.no_grad(): |
|
|
for batch in loader: |
|
|
image = batch["image"].to(device) |
|
|
case_id = batch["case_id"][0] if isinstance(batch["case_id"], (list, tuple)) else batch["case_id"] |
|
|
|
|
|
|
|
|
logits = sliding_window_inference( |
|
|
inputs=image, |
|
|
roi_size=tuple(cfg.infer.roi_size), |
|
|
sw_batch_size=cfg.infer.sw_batch_size, |
|
|
predictor=lambda x: model(x)[0], |
|
|
overlap=cfg.infer.overlap, |
|
|
) |
|
|
_, aux = model(image) |
|
|
probs = torch.sigmoid(logits) |
|
|
pi_et = aux["pi_et"].view(probs.shape[0], 1, 1, 1, 1) |
|
|
probs[:, 2:3] = probs[:, 2:3] * pi_et |
|
|
regions_bin = (probs > cfg.infer.threshold).float() |
|
|
|
|
|
|
|
|
et_pp = remove_small_components(regions_bin[:, 2], cfg.infer.et_cc_min_size) |
|
|
regions_bin[:, 2] = et_pp |
|
|
|
|
|
label_map = regions_to_label(regions_bin) |
|
|
|
|
|
meta = batch.get("image_meta_dict", None) |
|
|
affine = _get_affine(meta) |
|
|
prob_np = probs[0].detach().cpu().numpy().transpose(1, 2, 3, 0) |
|
|
bin_np = regions_bin[0].detach().cpu().numpy().transpose(1, 2, 3, 0) |
|
|
lbl_np = label_map[0, 0].detach().cpu().numpy().astype(np.int16) |
|
|
|
|
|
save_nifti(os.path.join(args.output, f"{case_id}_regions_prob.nii.gz"), prob_np, affine) |
|
|
save_nifti(os.path.join(args.output, f"{case_id}_regions_bin.nii.gz"), bin_np, affine) |
|
|
save_nifti(os.path.join(args.output, f"{case_id}_label.nii.gz"), lbl_np, affine) |
|
|
|
|
|
if data_format == "segmamba_npz": |
|
|
seg_path = os.path.join(args.output, f"{case_id}.nii.gz") |
|
|
seg_arr = regions_bin[0].detach().cpu().numpy().astype(np.uint8) |
|
|
save_segmamba_3c(seg_path, seg_arr, affine) |
|
|
|
|
|
|
|
|
if __name__ == "__main__": |
|
|
main() |
|
|
|
