Add GliomaSAM3-MoE and SegMamba model weights

.gitattributes

@@ -57,3 +57,12 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+gliomasam3_moe/vis_res/ampmix/Fig8_a_ampmix_BraTS-GLI-00005-000.pdf filter=lfs diff=lfs merge=lfs -text
+gliomasam3_moe/vis_res/boundary/Fig3_a_boundary_BraTS-GLI-00005-000.pdf filter=lfs diff=lfs merge=lfs -text
+gliomasam3_moe/vis_res/boundary/Fig3_b_boundary_BraTS-GLI-00017-000.pdf filter=lfs diff=lfs merge=lfs -text
+gliomasam3_moe/vis_res/concept_tokens/Fig6_b1_et_overview_BraTS-GLI-00006-000.pdf filter=lfs diff=lfs merge=lfs -text
+gliomasam3_moe/vis_res/dual_domain/Fig7_a_dual_domain_BraTS-GLI-00005-000.pdf filter=lfs diff=lfs merge=lfs -text
+gliomasam3_moe/vis_res/dual_domain/Fig7_b_dual_domain_BraTS-GLI-00017-000.pdf filter=lfs diff=lfs merge=lfs -text
+gliomasam3_moe/vis_res/et_absent/Fig2_a_et_absent_BraTS-GLI-00012-000.pdf filter=lfs diff=lfs merge=lfs -text
+gliomasam3_moe/vis_res/failure/Fig9_a_failure_BraTS-GLI-00020-000.pdf filter=lfs diff=lfs merge=lfs -text
+gliomasam3_moe/vis_res/qualitative/Fig1_qualitative_comparison.pdf filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,97 @@
+# SegMamba & GliomaSAM3-MoE Model Weights and Results
+
+This repository contains pre-trained model weights, evaluation results, and visualization outputs for brain tumor segmentation on BraTS 2023 dataset.
+
+## Repository Structure
+
+```
+.
+├── gliomasam3_moe/
+│   ├── checkpoints/           # GliomaSAM3-MoE model weights
+│   │   ├── ckpt_step2000.pt
+│   │   ├── ckpt_step2600.pt
+│   │   └── ckpt_step3000.pt   # Best checkpoint
+│   ├── configs/
+│   │   └── train.yaml         # Training configuration
+│   ├── eval_results/
+│   │   ├── table4_et_absent.json   # ET presence classification results
+│   │   └── table7_boundary_dice.json # Boundary-band Dice results
+│   └── vis_res/               # Visualization results
+│       ├── method_comparison/ # Side-by-side comparisons
+│       ├── boundary/          # Boundary analysis figures
+│       ├── moe_routing/       # MoE routing visualizations
+│       └── ...
+│
+├── segmamba/
+│   ├── checkpoints/           # SegMamba model weights
+│   │   ├── tmp_model_ep599_0.8295.pt
+│   │   └── tmp_model_ep799_0.8498.pt  # Best checkpoint (Dice=0.8498)
+│   └── prediction_results/
+│       ├── segmamba_brats23_ep799/    # Prediction NIfTI files
+│       └── result_metrics/            # Evaluation metrics
+│
+└── README.md
+```
+
+## Model Performance
+
+### GliomaSAM3-MoE (ckpt_step3000)
+
+**Boundary-band Dice (3-voxel band):**
+| Region | Dice |
+|--------|------|
+| WT | 0.789 ± 0.057 |
+| TC | 0.766 ± 0.154 |
+| ET | 0.697 ± 0.161 |
+| **Mean** | **0.750** |
+
+**ET Presence Classification:**
+| Metric | Value |
+|--------|-------|
+| AUROC | 0.896 |
+| Accuracy | 0.795 |
+| Sensitivity | 0.792 |
+| Specificity | 1.000 |
+
+### SegMamba (ep799)
+
+- Mean Dice: 0.8498
+- Trained for 800 epochs on BraTS 2023
+
+## Usage
+
+### Loading GliomaSAM3-MoE
+
+```python
+import torch
+
+# Load checkpoint
+ckpt = torch.load("gliomasam3_moe/checkpoints/ckpt_step3000.pt", map_location="cpu")
+
+# Model state dict is in ckpt["model"]
+model.load_state_dict(ckpt["model"])
+```
+
+### Loading SegMamba
+
+```python
+import torch
+
+# Load checkpoint
+ckpt = torch.load("segmamba/checkpoints/tmp_model_ep799_0.8498.pt", map_location="cpu")
+
+# Model state dict
+model.load_state_dict(ckpt["model"])
+```
+
+## Data
+
+Models were trained and evaluated on BraTS 2023 GLI Challenge dataset (122 cases in processed subset).
+
+## Citation
+
+If you use these models, please cite the relevant papers.
+
+## License
+
+Please refer to the original model repositories for licensing information.

gliomasam3_moe/checkpoints/ckpt_step2000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:886d389ca84282d411b5a8643d578bc540ef32477c05b498a66d2ef8c62e6507
+size 1821693123

gliomasam3_moe/checkpoints/ckpt_step2600.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5886858a63de35d717697158e747060ca9df3104a2d92637cb040966a8af96e2
+size 1821693123

gliomasam3_moe/checkpoints/ckpt_step3000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ffe568c3046807bb3b15bd36427ac935cd495acac15cd4937988b92f85cbab1a
+size 1821693123

gliomasam3_moe/configs/train.yaml

@@ -0,0 +1,76 @@
+seed: 42
+device: "cuda"
+amp: true
+synthetic: false
+
+data:
+  format: "segmamba_npz"
+  root_dir: "/data/yty/brats23_segmamba_processed"
+  modalities: ["t1n", "t1c", "t2f", "t2w"]
+  seg_name: "seg"
+  orientation: "RAS"
+  do_spacing: false
+  spacing: [1.0, 1.0, 1.0]
+  crop_size: [128, 128, 128]
+  num_samples: 1
+  rand_scale_prob: 0.1
+  rand_shift_prob: 0.1
+  batch_size: 6
+  synthetic_shape: [16, 128, 128]
+  synthetic_cases: 16
+  train_rate: 0.7
+  val_rate: 0.1
+  test_rate: 0.2
+  segmamba_unpack: true
+
+model:
+  patch_size: 14
+  token_dim: 128
+  depth: 3
+  heads: 4
+  mlp_ratio: 4.0
+  slice_attn_k: 6
+  slice_attn_random_dir: true
+  spectral_bins: 24
+  spectral_q: 3
+  msda_scales: [3, 5, 7]
+  moe_experts: 5
+  moe_topk: 2
+  decoder_hidden: 96
+  prompt_mlp_hidden: 128
+  use_sam3_backbone: true
+  sam3_ckpt_path: "/data/yty/sam3/sam3.pt"
+  sam3_freeze: true
+  sam3_in_chans: 7
+  sam3_input_mean: [0.5, 0.5, 0.5]
+  sam3_input_std: [0.5, 0.5, 0.5]
+
+loss:
+  dice_weight: 1.0
+  bce_weight: 1.0
+  et_focal_weight: 0.5
+  focal_gamma: 2.0
+  pres_weight: 0.1
+  hier_weight: 0.1
+  moe_weight: 0.01
+
+train:
+  epochs: 300
+  max_steps: 10000
+  lr: 0.0002
+  weight_decay: 0.00001
+  log_every: 20
+  save_every: 200
+  ckpt_dir: "./logs/segmamba/model"
+  fourier_mix_prob: 0.2
+  num_workers: 4
+  use_label_prompt: true
+  test_every_epochs: 5
+  test_max_cases: 0
+
+infer:
+  roi_size: [128, 128, 128]
+  sw_batch_size: 1
+  overlap: 0.5
+  threshold: 0.5
+  et_cc_min_size: 50

gliomasam3_moe/eval_results/table4_et_absent.json

@@ -0,0 +1,43 @@
+{
+  "et_absent_subset": {
+    "n": 2,
+    "fp_volume_mm3": {
+      "mean": 74.5,
+      "std": 74.5,
+      "min": 0.0,
+      "max": 149.0
+    },
+    "fp_components": {
+      "mean": 0.5,
+      "std": 0.5,
+      "min": 0,
+      "max": 1
+    }
+  },
+  "et_absent_cases": [
+    {
+      "case_id": "BraTS-GLI-00017-001",
+      "fp_volume_mm3": 149,
+      "fp_components": 1
+    },
+    {
+      "case_id": "BraTS-GLI-00048-001",
+      "fp_volume_mm3": 0,
+      "fp_components": 0
+    }
+  ],
+  "classification": {
+    "n": 122,
+    "n_et_present": 120,
+    "n_et_absent": 2,
+    "auroc": 0.8958333333333333,
+    "optimal_threshold": 0.9947388768196106,
+    "accuracy_optimal": 0.7950819672131147,
+    "sensitivity_optimal": 0.7916666666666666,
+    "specificity_optimal": 1.0
+  },
+  "config": {
+    "min_size": 50,
+    "threshold": 0.5
+  }
+}

gliomasam3_moe/eval_results/table7_boundary_dice.json

@@ -0,0 +1,26 @@
+{
+  "stats": {
+    "WT": {
+      "mean": 0.7889820841743286,
+      "std": 0.05687755328186895,
+      "n": 122
+    },
+    "TC": {
+      "mean": 0.7656578316169296,
+      "std": 0.1538429439512892,
+      "n": 122
+    },
+    "ET": {
+      "mean": 0.6964991356739063,
+      "std": 0.1612273062938224,
+      "n": 122
+    },
+    "Mean": {
+      "mean": 0.7503796838217216
+    }
+  },
+  "config": {
+    "radius": 3,
+    "min_size": 50
+  }
+}

gliomasam3_moe/vis_res/README.md

@@ -0,0 +1,206 @@
+# GliomaSAM3-MoE 可视化结果总结
+
+本文档总结了各可视化实验的内容、目的和主要结论。
+
+---
+
+## 1. 主定性对比 (qualitative/)
+
+**文件**: `Fig1_qualitative_comparison.png`
+
+**内容**: 
+- 多模态输入（T1, T1ce, T2, FLAIR）
+- Ground Truth 与预测结果对比
+- 我们的方法 (GliomaSAM3-MoE) vs SegMamba
+- WT/TC/ET 三区域分割叠加显示
+
+**结论**: 展示模型在不同病例上的定性分割效果，通过彩色叠加可以直观比较各方法在肿瘤边界、区域完整性上的差异。
+
+---
+
+## 2. ET Gate 研究 (et_absent/)
+
+**文件**: `Fig2_a_et_absent_BraTS-GLI-00012-000.png`
+
+**内容**:
+- ET Before Gate: 经过 gate 前的 ET 预测概率图
+- ET After Gate: 经过 gate 后的 ET 预测概率图
+- π_ET 值: 模型预测的 ET 存在概率
+
+**结论**: 
+- π_ET 值反映模型对 ET（增强肿瘤）存在性的判断置信度
+- 当 π_ET > 0.5 时，模型认为存在 ET；反之则抑制 ET 预测
+- 此机制可帮助减少 ET 假阳性，但在当前训练下效果较弱（大多数 π_ET 接近 0.97-0.98）
+
+---
+
+## 3. 边界误差分析 (boundary/)
+
+**文件**: 
+- `Fig3_a_boundary_BraTS-GLI-00005-000.png`
+- `Fig3_b_boundary_BraTS-GLI-00017-000.png`
+
+**内容**:
+- 左侧: T1ce 图像 + Ground Truth 边界（白色）+ 预测边界（黑色）
+- 右侧: 边界误差热力图（红色=假阳性FP，蓝色=假阴性FN）
+
+**结论**: 
+- 边界误差热力图直观展示分割误差的空间分布
+- 可用于分析 HD95 指标的改善区域
+- 误差主要集中在肿瘤边缘模糊区域
+
+---
+
+## 4. 微小/碎片 ET 分析 (tiny_et/)
+
+**文件**:
+- `Fig4_a_tiny_et_BraTS-GLI-00005-000.png`
+- `Fig4_b_tiny_et_BraTS-GLI-00006-000.png`
+
+**内容**:
+- 局部 ROI 放大显示
+- Ground Truth vs 我们的方法 vs SegMamba
+- 专注于 ET 极小或碎片化区域
+
+**结论**: 展示模型对微小 ET 区域的捕获能力，验证是否能保留碎片化结构而不丢失或过度平滑。
+
+---
+
+## 5. MoE 路由可解释性 (moe_routing/)
+
+**文件**:
+- `Fig5_a_moe_routing_BraTS-GLI-00018-000.png`
+- `Fig5_b_moe_routing_BraTS-GLI-00012-000.png`
+
+**内容**:
+- 柱状图展示各专家对 WT/TC/ET 三个区域的贡献权重
+- X 轴: 专家索引 (E0-E4)，激活的专家用 "(active)" 标注
+- Y 轴: 贡献值
+- 颜色区分: WT(青色), TC(品红), ET(黄色)
+- 黑色折线: 路由权重 γ，激活专家标注具体值
+- 非激活专家显示为半透明（alpha=0.3）
+
+**重要说明**:
+- 模型配置 `moe_topk: 2`，即采用 **Top-2 稀疏门控**
+- 每次推理**只激活 2 个专家**，其余专家权重为 0
+- 这是 Mixture-of-Experts 的标准设计，不是 bug
+
+**结论**:
+- Top-2 稀疏门控使计算效率提高，同时保持模型容量
+- 不同案例可能激活不同的专家组合（如 E0+E2 或其他）
+- 路由权重 γ 显示各激活专家的相对重要性
+
+---
+
+## 6. 概念 Token 可解释性 (concept_tokens/)
+
+每个病例生成 3 张图：
+
+### 6.1 ET 预测概览 (Fig6_X1_et_overview_*.png)
+- 左: T1ce 原始输入
+- 右: ET 预测 mask 叠加 + 体素总数
+
+### 6.2 碎片化分析 (Fig6_X2_fragmentation_*.png)
+- 左: 连通域可视化（不同颜色标识各组件）
+- 中: 各组件大小柱状图
+- 右: FRAG_BIN 分类（None/Low/Medium/High）
+
+### 6.3 规模分析 (Fig6_X3_scale_*.png)
+- 左: ET 区域可视化
+- 右: SCALE_BIN 分类（Tiny/Small/Medium/Large）
+
+**结论**:
+- **FRAG_BIN**: 根据 ET 连通域数量判断碎片化程度
+  - n ≤ 1 → None, n ≤ 3 → Low, n ≤ 5 → Medium, n > 5 → High
+- **SCALE_BIN**: 根据 ET 体素总数判断规模
+  - voxels ≤ 50 → Tiny, ≤ 200 → Small, ≤ 500 → Medium, > 500 → Large
+- 这些概念 token 与肿瘤形态特征直接关联，可用于辅助临床决策
+
+---
+
+## 7. 双域增强效果 (dual_domain/)
+
+**文件**:
+- `Fig7_a_dual_domain_BraTS-GLI-00005-000.png`
+- `Fig7_b_dual_domain_BraTS-GLI-00017-000.png`
+
+**内容**:
+- 左: 原始图像的傅里叶幅度谱
+- 右: 增强后的傅里叶幅度谱
+- 统一色标便于比较
+
+**结论**:
+- 频域可视化展示模型如何利用频谱信息
+- 增强后的幅度谱可能显示高频成分增强，有助于边界检测
+- 体现双域（空域+频域）融合的效果
+
+---
+
+## 8. AmpMix 增强鲁棒性 (ampmix/)
+
+**文件**: `Fig8_a_ampmix_BraTS-GLI-00005-000.png`
+
+**内容**:
+- 左: 原始图像 + 预测
+- 中: AmpMix 扰动后的图像
+- 右: 扰动后的预测结果
+
+**结论**:
+- AmpMix 通过混合不同样本的傅里叶幅度谱进行数据增强
+- 展示模型在幅度谱扰动下的预测稳定性
+- 理想情况下，扰动前后预测应保持一致，边界不发生显著漂移
+
+---
+
+## 9. 失败案例分析 (failure/)
+
+**文件**: `Fig9_a_failure_BraTS-GLI-00020-000.png`
+
+**内容**:
+- 左: Ground Truth
+- 右: 预测结果
+- 标注: 失败原因说明
+
+**结论**:
+- 展示模型的局限性
+- 典型失败原因包括:
+  - 边界模��区域的判断困难
+  - 低对比度区域的误分割
+  - 伪影或异常强度影响
+- 诚实展示模型限制有助于后续改进方向
+
+---
+
+## 文件结构
+
+```
+vis_res/
+├── README.md                 # 本文档
+├── qualitative/              # 主定性对比
+├── et_absent/                # ET Gate 研究
+├── boundary/                 # 边界误差分析
+├── tiny_et/                  # 微小 ET 分析
+├── moe_routing/              # MoE 路由可解释性
+├── concept_tokens/           # 概念 Token 可解释性
+├── dual_domain/              # 双域增强效果
+├── ampmix/                   # AmpMix 鲁棒性
+├── failure/                  # 失败案例
+└── aux_cache/                # 中间结果缓存（内部使用）
+```
+
+---
+
+## 生成方式
+
+所有可视化由 `visualizations/vis_publication.py` 生成，配置文件为 `visualizations/vis_config.yaml`。
+
+```bash
+cd /root/githubs/gliomasam3_moe
+PYTHONPATH=/root/githubs/sam3:$PYTHONPATH python visualizations/vis_publication.py \
+  --config visualizations/vis_config.yaml \
+  --model-config configs/train.yaml \
+  --checkpoint ./logs/segmamba/model/ckpt_step600.pt \
+  --run all
+```
+
+可通过 `--run <name>` 单独生成特定可视化（如 `--run moe,dual_domain`）。

gliomasam3_moe/vis_res/ampmix/Fig8_a_ampmix_BraTS-GLI-00005-000.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a4e155669c0f2ae5ba5340a49a6a131afdba8873a179db1efb67b7fdf3c022e1
+size 132112

gliomasam3_moe/vis_res/boundary/Fig3_a_boundary_BraTS-GLI-00005-000.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:410b3dddbb75fa28373d0fa9bb245bba0be27fa3e794ae23cb0c3b084cdec6bd
+size 128886

gliomasam3_moe/vis_res/boundary/Fig3_b_boundary_BraTS-GLI-00017-000.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:674e7fdabbf5d1cb909d03a46472aee1ee31c07b1b4126a5e508b3983269c74b
+size 132941

gliomasam3_moe/vis_res/concept_tokens/Fig6_b1_et_overview_BraTS-GLI-00006-000.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c89698cb5dae181364d13a93fda036899144c3e52494b58bc635c11d1a29fcd4
+size 102141

gliomasam3_moe/vis_res/dual_domain/Fig7_a_dual_domain_BraTS-GLI-00005-000.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fbae61c80f8bf576c539fae47daf8c65e84997288a4a636fad1b4dce637ce118
+size 113416

gliomasam3_moe/vis_res/dual_domain/Fig7_b_dual_domain_BraTS-GLI-00017-000.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:64d31767498b3c52967f9cd3e9edec59e97a0fa41a46d8767350dff97685316c
+size 131093

gliomasam3_moe/vis_res/et_absent/Fig2_a_et_absent_BraTS-GLI-00012-000.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1faf4ae3266d69db1efddae19f4ff18ea5b54333dcbe4587b4fe5859ec3e3d46
+size 144235

gliomasam3_moe/vis_res/failure/Fig9_a_failure_BraTS-GLI-00020-000.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:400eb6bd7ee79191f9a7dab3825d65aefdbfbd3495be2cc8bdaffa7fee3856de
+size 138511