source_code/sam3/medsam3_brats/preprocess_brats.py

#!/usr/bin/env python3
"""
BraTS2023 数据预处理脚本
将3D NIfTI医学数据转换为SAM3可处理的帧序列格式
"""

import os
import argparse
import numpy as np
import nibabel as nib
from pathlib import Path
from tqdm import tqdm
import cv2
from PIL import Image
import json
import multiprocessing as mp


def normalize_intensity(volume, low_percentile=0.5, high_percentile=99.5):
    """
    对体积数据进行强度归一化
    """
    low = np.percentile(volume[volume > 0], low_percentile)
    high = np.percentile(volume[volume > 0], high_percentile)
    volume = np.clip(volume, low, high)
    volume = (volume - low) / (high - low + 1e-8)
    return volume


def load_brats_case(case_dir):
    """
    加载单个BraTS病例的所有模态
    
    Args:
        case_dir: 病例文件夹路径
    
    Returns:
        data: 形状为 (4, D, H, W) 的numpy数组，4个模态
        seg: 形状为 (D, H, W) 的分割标签
        affine: NIfTI仿射矩阵
    """
    case_dir = Path(case_dir)
    case_name = case_dir.name
    
    # BraTS2023 模态
    modalities = ['t1c', 't1n', 't2f', 't2w']
    
    data = []
    affine = None
    
    for mod in modalities:
        # 尝试不同的命名格式
        possible_names = [
            f"{mod}.nii.gz",
            f"{case_name}-{mod}.nii.gz",
            f"{case_name}_{mod}.nii.gz"
        ]
        
        nii_path = None
        for name in possible_names:
            p = case_dir / name
            if p.exists():
                nii_path = p
                break
        
        if nii_path is None:
            raise FileNotFoundError(f"Cannot find {mod} file in {case_dir}")
        
        nii = nib.load(str(nii_path))
        if affine is None:
            affine = nii.affine
        
        volume = nii.get_fdata().astype(np.float32)
        volume = normalize_intensity(volume)
        data.append(volume)
    
    data = np.stack(data, axis=0)  # (4, D, H, W)
    
    # 加载分割标签
    seg_names = [
        "seg.nii.gz",
        f"{case_name}-seg.nii.gz",
        f"{case_name}_seg.nii.gz"
    ]
    
    seg = None
    for name in seg_names:
        p = case_dir / name
        if p.exists():
            seg_nii = nib.load(str(p))
            seg = seg_nii.get_fdata().astype(np.int32)
            break
    
    return data, seg, affine


def convert_to_frames(data, output_dir, case_name, modality_idx=0, target_size=(512, 512)):
    """
    将3D数据转换为帧序列（JPEG图像）
    
    Args:
        data: 形状为 (4, D, H, W) 的数据
        output_dir: 输出目录
        case_name: 病例名称
        modality_idx: 使用哪个模态 (0=t1c, 1=t1n, 2=t2f, 3=t2w)
        target_size: 目标图像大小
    """
    frames_dir = Path(output_dir) / case_name / "frames"
    frames_dir.mkdir(parents=True, exist_ok=True)
    
    # 选择模态
    volume = data[modality_idx]  # (D, H, W)
    
    # 转换为 (D, H, W) 格式，D为深度（切片数）
    # BraTS数据通常是 (H, W, D)，需要转置
    if volume.shape[0] > volume.shape[2]:
        volume = np.transpose(volume, (2, 0, 1))
    
    num_slices = volume.shape[0]
    
    for i in range(num_slices):
        slice_2d = volume[i]  # (H, W)
        
        # 归一化到 0-255
        slice_2d = (slice_2d * 255).astype(np.uint8)
        
        # 转为RGB（SAM3需要RGB输入）
        slice_rgb = np.stack([slice_2d, slice_2d, slice_2d], axis=-1)
        
        # 调整大小
        if target_size is not None:
            slice_rgb = cv2.resize(slice_rgb, target_size, interpolation=cv2.INTER_LINEAR)
        
        # 保存为JPEG
        frame_path = frames_dir / f"{i:05d}.jpg"
        Image.fromarray(slice_rgb).save(str(frame_path), quality=95)
    
    return frames_dir, num_slices


def save_segmentation_masks(seg, output_dir, case_name, target_size=(512, 512)):
    """
    保存分割标签为帧序列
    
    BraTS标签:
        0: 背景
        1: NCR (Necrotic Core) - 坏死核心
        2: ED (Edema) - 水肿
        3: ET (Enhancing Tumor) - 强化肿瘤
        
    合并为:
        - Whole Tumor (WT): 1+2+3
        - Tumor Core (TC): 1+3
        - Enhancing Tumor (ET): 3
    """
    if seg is None:
        return None
    
    masks_dir = Path(output_dir) / case_name / "masks"
    masks_dir.mkdir(parents=True, exist_ok=True)
    
    # 转置如果需要
    if seg.shape[0] > seg.shape[2]:
        seg = np.transpose(seg, (2, 0, 1))
    
    num_slices = seg.shape[0]
    
    # 创建不同区域的mask
    for i in range(num_slices):
        slice_seg = seg[i]  # (H, W)
        
        # Whole Tumor (包含所有肿瘤区域)
        wt_mask = ((slice_seg == 1) | (slice_seg == 2) | (slice_seg == 3)).astype(np.uint8) * 255
        
        # 调整大小
        if target_size is not None:
            wt_mask = cv2.resize(wt_mask, target_size, interpolation=cv2.INTER_NEAREST)
        
        # 保存mask
        mask_path = masks_dir / f"{i:05d}.png"
        Image.fromarray(wt_mask).save(str(mask_path))
    
    return masks_dir


def get_tumor_bbox_and_center(seg, slice_idx=None):
    """
    获取肿瘤的边界框和中心点
    
    Args:
        seg: 分割标签 (D, H, W)
        slice_idx: 指定切片索引，如果为None则找最大肿瘤面积的切片
    
    Returns:
        slice_idx: 选中的切片索引
        bbox: (x_min, y_min, x_max, y_max)
        center: (x, y)
    """
    if seg is None:
        return None, None, None
    
    # 转置如果需要
    if seg.shape[0] > seg.shape[2]:
        seg = np.transpose(seg, (2, 0, 1))
    
    # 整个肿瘤区域
    tumor_mask = (seg > 0)
    
    # 如果没有指定切片，找最大肿瘤面积的切片
    if slice_idx is None:
        tumor_areas = tumor_mask.sum(axis=(1, 2))
        slice_idx = int(np.argmax(tumor_areas))
    
    # 获取该切片的肿瘤mask
    slice_tumor = tumor_mask[slice_idx]
    
    if slice_tumor.sum() == 0:
        return slice_idx, None, None
    
    # 找边界框
    rows = np.any(slice_tumor, axis=1)
    cols = np.any(slice_tumor, axis=0)
    y_min, y_max = np.where(rows)[0][[0, -1]]
    x_min, x_max = np.where(cols)[0][[0, -1]]
    
    bbox = (int(x_min), int(y_min), int(x_max), int(y_max))
    center = ((x_min + x_max) // 2, (y_min + y_max) // 2)
    
    return slice_idx, bbox, center


def save_prompt_info(output_dir, case_name, slice_idx, bbox, center, 
                     original_size, target_size=(512, 512)):
    """
    保存提示信息（用于SAM3推理）
    """
    info_dir = Path(output_dir) / case_name
    info_dir.mkdir(parents=True, exist_ok=True)
    
    # 计算缩放比例
    scale_x = target_size[0] / original_size[1]
    scale_y = target_size[1] / original_size[0]
    
    # 缩放bbox和center
    if bbox is not None:
        scaled_bbox = (
            int(bbox[0] * scale_x),
            int(bbox[1] * scale_y),
            int(bbox[2] * scale_x),
            int(bbox[3] * scale_y)
        )
    else:
        scaled_bbox = None
    
    if center is not None:
        scaled_center = (
            int(center[0] * scale_x),
            int(center[1] * scale_y)
        )
    else:
        scaled_center = None
    
    info = {
        'slice_idx': slice_idx,
        'bbox': scaled_bbox,
        'center': scaled_center,
        'original_size': original_size,
        'target_size': target_size,
        'scale': (scale_x, scale_y)
    }
    
    np.save(str(info_dir / 'prompt_info.npy'), info, allow_pickle=True)
    
    return info


def process_single_case(case_dir, output_dir, modality_idx=0, target_size=(512, 512)):
    """
    处理单个病例
    """
    case_name = Path(case_dir).name
    print(f"Processing {case_name}...")
    
    try:
        # 加载数据
        data, seg, affine = load_brats_case(case_dir)
        original_size = data.shape[2:4]  # (H, W)
        
        # 转换为帧
        frames_dir, num_slices = convert_to_frames(
            data, output_dir, case_name, 
            modality_idx=modality_idx, 
            target_size=target_size
        )
        
        # 保存mask
        masks_dir = save_segmentation_masks(seg, output_dir, case_name, target_size=target_size)
        
        # 获取提示信息
        slice_idx, bbox, center = get_tumor_bbox_and_center(seg)
        
        # 保存提示信息
        prompt_info = save_prompt_info(
            output_dir, case_name, slice_idx, bbox, center,
            original_size, target_size
        )
        
        print(f"  - {num_slices} slices processed")
        print(f"  - Tumor center slice: {slice_idx}")
        if bbox:
            print(f"  - BBox: {prompt_info['bbox']}")
            print(f"  - Center: {prompt_info['center']}")
        
        return True
        
    except Exception as e:
        print(f"Error processing {case_name}: {e}")
        import traceback
        traceback.print_exc()
        return False


def _make_splits(case_names, train_ratio: float, val_ratio: float, test_ratio: float, seed: int):
    """Deterministic case-level split into train/val/test."""
    if not (0 < train_ratio < 1) or not (0 <= val_ratio < 1) or not (0 <= test_ratio < 1):
        raise ValueError("ratios must be in [0,1) and train_ratio in (0,1)")
    if abs((train_ratio + val_ratio + test_ratio) - 1.0) > 1e-6:
        raise ValueError(f"train/val/test ratios must sum to 1.0, got {train_ratio+val_ratio+test_ratio:.6f}")

    case_names = list(case_names)
    rng = np.random.RandomState(seed)
    rng.shuffle(case_names)

    n = len(case_names)
    n_train = int(round(n * train_ratio))
    n_val = int(round(n * val_ratio))
    # ensure non-negative and all used
    n_train = min(max(n_train, 0), n)
    n_val = min(max(n_val, 0), n - n_train)
    n_test = n - n_train - n_val

    train = case_names[:n_train]
    val = case_names[n_train : n_train + n_val]
    test = case_names[n_train + n_val :]
    return {"train": train, "val": val, "test": test}


def _already_processed(output_dir: Path, case_name: str) -> bool:
    """Heuristic for resuming: prompt_info.npy exists AND frames/masks dirs exist."""
    case_dir = output_dir / case_name
    if not (case_dir / "prompt_info.npy").exists():
        return False
    if not (case_dir / "frames").is_dir():
        return False
    if not (case_dir / "masks").is_dir():
        return False
    return True


def _worker_process_case(args):
    case_dir, output_dir, modality_idx, target_size, skip_existing = args
    case_name = Path(case_dir).name
    output_dir = Path(output_dir)
    if skip_existing and _already_processed(output_dir, case_name):
        return case_name, True, "skipped"
    ok = process_single_case(case_dir, output_dir, modality_idx=modality_idx, target_size=target_size)
    return case_name, ok, "processed" if ok else "failed"


def main():
    parser = argparse.ArgumentParser(description='Preprocess BraTS data for SAM3')
    parser.add_argument('--input_dir', type=str, required=True,
                        help='Input directory containing BraTS cases')
    parser.add_argument('--output_dir', type=str, required=True,
                        help='Output directory for processed data')
    parser.add_argument('--modality', type=int, default=0,
                        help='Modality index: 0=t1c, 1=t1n, 2=t2f, 3=t2w')
    parser.add_argument('--target_size', type=int, nargs=2, default=[512, 512],
                        help='Target image size (width height)')
    parser.add_argument('--num_cases', type=int, default=None,
                        help='Number of cases to process (None for all)')
    parser.add_argument('--num_processes', type=int, default=1,
                        help='Number of parallel worker processes for preprocessing')
    parser.add_argument('--skip_existing', action='store_true',
                        help='Skip cases that already have frames/masks/prompt_info.npy in output_dir')

    # split file generation (case-level)
    parser.add_argument('--write_splits', action='store_true',
                        help='Write train/val/test split json to output_dir/splits.json')
    parser.add_argument('--train_ratio', type=float, default=0.7)
    parser.add_argument('--val_ratio', type=float, default=0.1)
    parser.add_argument('--test_ratio', type=float, default=0.2)
    parser.add_argument('--seed', type=int, default=42)
    
    args = parser.parse_args()
    
    input_dir = Path(args.input_dir)
    output_dir = Path(args.output_dir)
    output_dir.mkdir(parents=True, exist_ok=True)
    
    # 获取所有病例
    case_dirs = sorted([d for d in input_dir.iterdir() if d.is_dir()])
    
    if args.num_cases is not None:
        case_dirs = case_dirs[:args.num_cases]
    
    print(f"Found {len(case_dirs)} cases to process")

    # write split file (based on all available case dirs)
    if args.write_splits:
        splits = _make_splits(
            [Path(d).name for d in case_dirs],
            train_ratio=args.train_ratio,
            val_ratio=args.val_ratio,
            test_ratio=args.test_ratio,
            seed=args.seed,
        )
        split_path = output_dir / "splits.json"
        with open(split_path, "w") as f:
            json.dump(
                {
                    "seed": args.seed,
                    "train_ratio": args.train_ratio,
                    "val_ratio": args.val_ratio,
                    "test_ratio": args.test_ratio,
                    "splits": splits,
                },
                f,
                indent=2,
            )
        print(
            f"Wrote splits to {split_path} "
            f"(train={len(splits['train'])}, val={len(splits['val'])}, test={len(splits['test'])})"
        )
    
    success_count = 0
    target_size = tuple(args.target_size)

    if args.num_processes <= 1:
        for case_dir in tqdm(case_dirs, desc="Processing cases"):
            case_name = Path(case_dir).name
            if args.skip_existing and _already_processed(output_dir, case_name):
                continue
            if process_single_case(case_dir, output_dir, modality_idx=args.modality, target_size=target_size):
                success_count += 1
    else:
        work = [
            (str(case_dir), str(output_dir), int(args.modality), target_size, bool(args.skip_existing))
            for case_dir in case_dirs
        ]
        with mp.get_context("spawn").Pool(processes=int(args.num_processes)) as pool:
            for case_name, ok, status in tqdm(
                pool.imap_unordered(_worker_process_case, work),
                total=len(work),
                desc="Processing cases (mp)",
            ):
                if ok and status != "skipped":
                    success_count += 1
    
    print(f"\nProcessed {success_count}/{len(case_dirs)} cases successfully")


if __name__ == "__main__":
    main()