1# 1. Clone repository
 2git clone https://github.com/xmed-lab/SD-Messenger.git
 3cd SD-Messenger/SD_Messenger
 4
 5# 2. 建立 conda 環境（建議用 mamba 加速）
 6conda create -n SDMessenger python=3.10 -y
 7conda activate SDMessenger
 8
 9# 3. 安裝 PyTorch（依據 CUDA 版本調整）
10pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
11
12# 4. 安裝 mmcv（需要對應 PyTorch + CUDA 版本）
13pip install mmcv-full==1.6.2 -f https://download.openmmlab.com/mmcv/dist/cu118/torch2.0/index.html
14
15# 5. 安裝其餘依賴
16pip install -r requirements.txt

 1# 下載預處理資料集（以 Synapse 為例）
 2# 從 OneDrive 連結下載 synapse.zip
 3
 4# 解壓縮
 5unzip synapse.zip -d ./data/
 6
 7# 目錄結構應為：
 8# SD_Messenger/
 9# ├── data/
10# │   └── synapse/
11# │       ├── train/
12# │       │   ├── images/
13# │       │   └── masks/
14# │       └── val/
15# │           ├── images/
16# │           └── masks/
17# ├── splits/
18# │   └── synapse/
19# │       ├── 1_5/          # 20% labeled
20# │       │   ├── labeled.txt
21# │       │   └── unlabeled.txt
22# │       ├── 2_5/          # 40% labeled
23# │       └── val.txt
24# └── configs/

1data_root: ./data/synapse

1cd SD_Messenger
2
3# 單 GPU 訓練
4bash scripts/train_synapse_1_5.sh 1 12345
5
6# 多 GPU 訓練（例如 4 張 GPU）
7bash scripts/train_synapse_1_5.sh 4 12345

 1#!/bin/bash
 2dataset='synapse'
 3method='train_synapse'
 4config='synapse_1_5'
 5split='1_5'
 6now=$(date +"%Y%m%d_%H%M%S")
 7config=configs/${config}.yaml
 8labeled_id_path=splits/$dataset/$split/labeled.txt
 9unlabeled_id_path=splits/$dataset/$split/unlabeled.txt
10save_path=exp/$method/$dataset/$split
11
12mkdir -p $save_path
13
14# torchrun 啟動分散式訓練
15CUDA_VISIBLE_DEVICES=0 torchrun \
16    --nproc_per_node=$1 \
17    --master_addr=localhost \
18    --master_port=$2 \
19    $method.py \
20    --config=$config \
21    --labeled-id-path $labeled_id_path \
22    --unlabeled-id-path $unlabeled_id_path \
23    --save-path $save_path \
24    --port $2 2>&1 | tee $save_path/$now.log

 1# 資料集設定
 2dataset: synapse
 3nclass: 14            # 14 個器官類別
 4crop_size: 512        # 輸入影像裁切尺寸
 5l2u_size: 64          # Labeled-to-Unlabeled 特徵圖尺寸
 6
 7# 訓練設定
 8epochs: 300
 9batch_size: 8         # 每張 GPU
10lr: 0.001
11lr_multi: 5.0         # 非 backbone 參數的學習率倍率
12
13# 模型設定
14model:
15  backbone:
16    type: model.backbone.mit.mit_b5    # MiT-B5 backbone
17    kwargs:
18      embed_dims: [64, 128, 320, 512]
19      pretrained: True
20  framework:
21    type: model.semseg.sd_messenger.SDMessenger
22    kwargs:
23      num_layers: 2
24      num_heads: 2
25      num_class: 14
26      in_planes: [64, 128, 320, 512]
27      image_size: 512
28      add_cross_attn: [False, False, False, True]  # 僅在最深層啟用

 1cd SD_Messenger
 2
 3# 使用預訓練權重評估 Synapse 20% 設定
 4bash scripts/test_synapse_1_5.sh 1 /path/to/checkpoint.pth
 5
 6# 或手動執行（更靈活的參數控制）
 7CUDA_VISIBLE_DEVICES=0 torchrun \
 8    --nproc_per_node=1 \
 9    --master_addr=localhost \
10    --master_port=12345 \
11    test_synapse.py \
12    --config=configs/synapse_1_5.yaml \
13    --checkpoint-path=exp/train_synapse/synapse/1_5/best_model.pth \
14    --port 12345

 1import torch
 2import torch.nn as nn
 3from model.semseg.sd_messenger import U2LModule, SDMessenger
 4
 5# --- 方式 A：使用完整 SDMessenger 框架 ---
 6sd_model = SDMessenger(
 7    num_layers=2,
 8    num_heads=2,
 9    num_class=5,                     # 自定義類別數
10    in_planes=[64, 128, 320, 512],   # 需匹配 backbone 輸出維度
11    image_size=256,                  # 輸入影像尺寸
12    add_cross_attn=[False, False, True, True]  # 在後兩層啟用
13)
14
15# 模擬 4 階段 backbone 輸出（batch = labeled + unlabeled）
16batch_size = 4  # 2 labeled + 2 unlabeled
17c1 = torch.randn(batch_size, 64, 64, 64)
18c2 = torch.randn(batch_size, 128, 32, 32)
19c3 = torch.randn(batch_size, 320, 16, 16)
20c4 = torch.randn(batch_size, 512, 8, 8)
21
22out, e3_enhanced = sd_model((c1, c2, c3, c4), h=256, w=256)
23print(f"Segmentation output: {out.shape}")       # [4, 5, 256, 256]
24print(f"Enhanced features:   {e3_enhanced.shape}") # [4, 512, 8, 8]
25
26# 分離 labeled / unlabeled 預測
27pred_labeled = out[:batch_size // 2]     # 用 ground-truth 監督
28pred_unlabeled = out[batch_size // 2:]   # 用 pseudo-label 監督
29
30# --- 方式 B：單獨使用 U2L Module ---
31u2l = U2LModule(
32    num_layers=2,
33    num_heads=2,
34    embedding_channels=512,
35    channel_num=512,
36    channel_num_out=512,
37    patchSize=1,
38    scale=1,
39    dropout_rate=0.5,
40    attention_dropout_rate=0.1,
41    num_class=5,
42    patch_num=65  # (image_size // 32 + 1) ** 2
43)
44
45# 輸入需為 labeled + unlabeled concat
46feat = torch.randn(4, 512, 8, 8)  # 2 labeled + 2 unlabeled
47enhanced = u2l(feat)
48print(f"U2L enhanced: {enhanced.shape}")  # [4, 512, 8, 8]
49# enhanced[:2] = domain-enriched labeled features
50# enhanced[2:] = self-attention enhanced unlabeled features

 1# 使用 MONAI 的資料載入 + S&D Messenger 的模型
 2import monai
 3from monai.transforms import (
 4    Compose, LoadImaged, EnsureChannelFirstd,
 5    ScaleIntensityd, RandCropByPosNegLabeld
 6)
 7from model.model_helper import ModelBuilder
 8
 9# MONAI 資料管線
10train_transforms = Compose([
11    LoadImaged(keys=["image", "label"]),
12    EnsureChannelFirstd(keys=["image", "label"]),
13    ScaleIntensityd(keys=["image"]),
14    RandCropByPosNegLabeld(
15        keys=["image", "label"],
16        label_key="label",
17        spatial_size=(512, 512),
18        pos=1, neg=1,
19    ),
20])
21
22# S&D Messenger 模型設定
23cfg_model = {
24    "backbone": {
25        "type": "model.backbone.mit.mit_b5",
26        "kwargs": {
27            "embed_dims": [64, 128, 320, 512],
28            "pretrained": True
29        }
30    },
31    "framework": {
32        "type": "model.semseg.sd_messenger.SDMessenger",
33        "kwargs": {
34            "num_layers": 2,
35            "num_heads": 2,
36            "num_class": 14,
37            "in_planes": [64, 128, 320, 512],
38            "image_size": 512,
39            "add_cross_attn": [False, False, False, True]
40        }
41    }
42}
43
44model = ModelBuilder(cfg_model)

 1# 1. 使用 nnUNet 做全監督 baseline（100% labeled）
 2nnUNetv2_plan_and_preprocess -d DATASET_ID --verify_dataset_integrity
 3nnUNetv2_train DATASET_ID 2d 0
 4
 5# 2. 使用 S&D Messenger 做半監督實驗（20% labeled）
 6cd SD_Messenger
 7bash scripts/train_synapse_1_5.sh 1 12345
 8
 9# 3. 比較 Dice 分數，評估 annotation efficiency
10# nnUNet 100% labeled vs S&D Messenger 20% labeled
11# 若 S&D Messenger 20% 逼近 nnUNet 100%，則標註效率提升 5 倍