File size: 21,783 Bytes
fe8202e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
"""
Ablation Evaluation Script for GliomaSAM3-MoE

Implements:
- Table 4: ET-absent subset evaluation
- Table 7: Boundary-band Dice (3-voxel band)

Usage:
    cd /root/githubs/gliomasam3_moe
    PYTHONPATH=/root/githubs/sam3:$PYTHONPATH python eval_ablation.py \
        --config configs/train.yaml \
        --checkpoint logs/segmamba/model/ckpt_step3000.pt \
        --eval table7  # or table4, or both

Author: GliomaSAM3-MoE Team
"""

import argparse
import os
import sys
import json
from typing import Dict, List, Tuple, Optional
from collections import defaultdict

import numpy as np
import torch
import yaml
from tqdm import tqdm

# Add project paths
ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
SRC_DIR = os.path.join(ROOT_DIR, "src")
if SRC_DIR not in sys.path:
    sys.path.insert(0, SRC_DIR)

from scipy import ndimage as ndi
from scipy.ndimage import binary_dilation, binary_erosion
from sklearn.metrics import roc_auc_score, accuracy_score, roc_curve

# ============================================================================
# Configuration
# ============================================================================
DEFAULT_CONFIG = {
    "data_dir": "/data/yty/brats23_segmamba_processed",
    "seed": 20251225,  # Fixed seed as per spec
    "train_rate": 0.7,
    "val_rate": 0.1,
    "test_rate": 0.2,
    "threshold": 0.5,
    "et_cc_min_size": 50,
    "boundary_band_radius": 3,
    "hd95_empty_value": 50.0,
}


# ============================================================================
# Utility Functions
# ============================================================================
def load_yaml(path: str) -> Dict:
    with open(path, "r") as f:
        return yaml.safe_load(f)


def split_npz_paths(data_dir: str, train_rate: float, val_rate: float, 
                    test_rate: float, seed: int) -> Tuple[List[str], List[str], List[str]]:
    """Split data paths into train/val/test sets with fixed seed."""
    import glob
    import random
    
    all_paths = sorted(glob.glob(os.path.join(data_dir, "*.npz")))
    
    random.seed(seed)
    random.shuffle(all_paths)
    
    n = len(all_paths)
    n_train = int(n * train_rate)
    n_val = int(n * val_rate)
    
    train_paths = all_paths[:n_train]
    val_paths = all_paths[n_train:n_train + n_val]
    test_paths = all_paths[n_train + n_val:]
    
    return train_paths, val_paths, test_paths


def load_case(npz_path: str) -> Dict:
    """Load a single case from npz/npy files."""
    npy_path = npz_path[:-4] + ".npy"
    seg_path = npz_path[:-4] + "_seg.npy"
    
    # Load image
    if os.path.isfile(npy_path):
        image = np.load(npy_path, mmap_mode="r")
    else:
        data = np.load(npz_path)
        image = data["data"]
    
    image = np.asarray(image, dtype=np.float32)
    if image.ndim == 5 and image.shape[0] == 1:
        image = image[0]
    if image.ndim == 4 and image.shape[0] != 4 and image.shape[-1] == 4:
        image = image.transpose(3, 0, 1, 2)
    
    # Load label
    if os.path.isfile(seg_path):
        label = np.load(seg_path, mmap_mode="r")
    else:
        data = np.load(npz_path)
        label = data["seg"] if "seg" in data else None
    
    if label is not None:
        label = np.asarray(label, dtype=np.int16)
        if label.ndim == 4 and label.shape[0] == 1:
            label = label[0]
        # Map ET label 3 -> 4 if needed (BraTS convention)
        if label.max() == 3 and (label == 4).sum() == 0:
            label = label.copy()
            label[label == 3] = 4
    
    case_id = os.path.basename(npz_path)[:-4]
    return {"image": image, "label": label, "case_id": case_id}


def label_to_regions(label: np.ndarray) -> np.ndarray:
    """Convert BraTS label to [WT, TC, ET] regions."""
    label = np.asarray(label)
    wt = label > 0
    tc = (label == 1) | (label == 4)
    et = label == 4
    return np.stack([wt, tc, et], axis=0).astype(np.uint8)


def remove_small_components(mask: np.ndarray, min_size: int, connectivity: int = 3) -> np.ndarray:
    """Remove connected components smaller than min_size.
    
    Args:
        mask: Binary mask [D, H, W]
        min_size: Minimum voxel count to keep
        connectivity: 1 for 6-connectivity, 2 for 18, 3 for 26
    """
    struct = ndi.generate_binary_structure(3, connectivity)
    labeled, num = ndi.label(mask.astype(np.uint8), structure=struct)
    if num == 0:
        return mask.astype(np.uint8)
    
    sizes = ndi.sum(mask.astype(np.uint8), labeled, index=np.arange(1, num + 1))
    keep = np.zeros_like(mask, dtype=np.uint8)
    for i, s in enumerate(sizes, start=1):
        if s >= min_size:
            keep[labeled == i] = 1
    return keep


def count_connected_components(mask: np.ndarray, connectivity: int = 3) -> int:
    """Count number of connected components."""
    struct = ndi.generate_binary_structure(3, connectivity)
    _, num = ndi.label(mask.astype(np.uint8), structure=struct)
    return num


# ============================================================================
# Model Inference
# ============================================================================
class ModelPredictor:
    """Predictor for GliomaSAM3-MoE model."""
    
    def __init__(self, config_path: str, checkpoint_path: str, device: str = "cuda"):
        self.device = torch.device(device if torch.cuda.is_available() else "cpu")
        self.cfg = load_yaml(config_path)
        
        from gliomasam3_moe.models.gliomasam3_moe import GliomaSAM3_MoE
        
        self.model = GliomaSAM3_MoE(**self.cfg["model"]).to(self.device)
        ckpt = torch.load(checkpoint_path, map_location="cpu")
        state_dict = {k: v for k, v in ckpt["model"].items() if "freqs_cis" not in k}
        self.model.load_state_dict(state_dict, strict=False)
        self.model.eval()
        print(f"Loaded checkpoint: {checkpoint_path}")
    
    def predict(self, image: np.ndarray) -> Dict:
        """Run inference and return predictions with aux outputs.
        
        Returns:
            {
                "probs": np.ndarray [3, D, H, W] - raw probabilities before gating
                "probs_gated": np.ndarray [3, D, H, W] - probabilities after ET gating
                "pi_et": float - ET presence probability
                "regions_bin": np.ndarray [3, D, H, W] - binary predictions (after gating + threshold)
            }
        """
        if image.ndim == 4:
            x = torch.from_numpy(image.copy()).float().unsqueeze(0)
        else:
            raise ValueError(f"Invalid image shape: {image.shape}")
        
        x = x.to(self.device)
        
        with torch.no_grad():
            logits, aux = self.model(x)
            probs = torch.sigmoid(logits)
            
            # Get pi_et
            pi_et = aux["pi_et"]
            pi_et_value = float(pi_et.item())
            
            # Probs before gating
            probs_np = probs[0].cpu().numpy()
            
            # Apply ET gating
            probs_gated = probs.clone()
            probs_gated[:, 2:3] = probs[:, 2:3] * pi_et.view(-1, 1, 1, 1, 1)
            probs_gated_np = probs_gated[0].cpu().numpy()
            
            # Binary prediction with threshold
            threshold = self.cfg.get("infer", {}).get("threshold", 0.5)
            regions_bin = (probs_gated_np > threshold).astype(np.uint8)
        
        return {
            "probs": probs_np,
            "probs_gated": probs_gated_np,
            "pi_et": pi_et_value,
            "regions_bin": regions_bin,
        }


# ============================================================================
# Table 7: Boundary-band Dice
# ============================================================================
def compute_boundary_band(mask: np.ndarray, radius: int = 3) -> np.ndarray:
    """Compute 3D boundary band using morphological operations.
    
    Args:
        mask: Binary mask [D, H, W]
        radius: Dilation/erosion radius in voxels
        
    Returns:
        band: Binary mask of boundary band
    """
    struct = ndi.generate_binary_structure(3, 3)  # 26-connectivity
    
    mask_bool = mask.astype(bool)
    if mask_bool.sum() == 0:
        return np.zeros_like(mask, dtype=np.uint8)
    
    # Dilate and erode
    dilated = binary_dilation(mask_bool, structure=struct, iterations=radius)
    eroded = binary_erosion(mask_bool, structure=struct, iterations=radius)
    
    # Band = dilated XOR eroded
    band = np.logical_xor(dilated, eroded).astype(np.uint8)
    return band


def compute_boundary_band_dice(pred: np.ndarray, gt: np.ndarray, radius: int = 3) -> float:
    """Compute Boundary-band Dice score.
    
    Args:
        pred: Binary prediction [D, H, W]
        gt: Binary ground truth [D, H, W]
        radius: Band radius in voxels
        
    Returns:
        Dice score for boundary band region
    """
    eps = 1e-7
    
    pred = pred.astype(bool)
    gt = gt.astype(bool)
    
    # Handle empty cases
    if gt.sum() == 0 and pred.sum() == 0:
        return 1.0
    if gt.sum() == 0 and pred.sum() > 0:
        return 0.0
    
    # Compute boundary band from GT
    band = compute_boundary_band(gt, radius=radius)
    
    # Restrict pred and gt to band
    pred_band = pred & band.astype(bool)
    gt_band = gt & band.astype(bool)
    
    # Dice on band
    intersection = (pred_band & gt_band).sum()
    dice = 2 * intersection / (pred_band.sum() + gt_band.sum() + eps)
    
    return float(dice)


def eval_table7(predictor: ModelPredictor, val_paths: List[str], 
                config: Dict, output_dir: str) -> Dict:
    """Evaluate Table 7: Boundary-band Dice (3-voxel band).
    
    Returns:
        Results dict with per-region and mean boundary dice
    """
    print("\n" + "=" * 60)
    print("Table 7: Boundary-band Dice Evaluation")
    print("=" * 60)
    
    radius = config.get("boundary_band_radius", 3)
    min_size = config.get("et_cc_min_size", 50)
    
    results = {
        "WT": [], "TC": [], "ET": [],
        "config": {"radius": radius, "min_size": min_size}
    }
    
    for npz_path in tqdm(val_paths, desc="Evaluating"):
        case = load_case(npz_path)
        if case["label"] is None:
            continue
        
        # Get predictions
        pred_out = predictor.predict(case["image"])
        pred_regions = pred_out["regions_bin"].copy()
        
        # Post-process ET: remove small components
        pred_regions[2] = remove_small_components(pred_regions[2], min_size, connectivity=3)
        
        # Get GT regions
        gt_regions = label_to_regions(case["label"])
        
        # Compute boundary-band dice for each region
        for i, region in enumerate(["WT", "TC", "ET"]):
            dice = compute_boundary_band_dice(pred_regions[i], gt_regions[i], radius=radius)
            results[region].append(dice)
    
    # Compute statistics
    stats = {}
    for region in ["WT", "TC", "ET"]:
        scores = results[region]
        stats[region] = {
            "mean": float(np.mean(scores)),
            "std": float(np.std(scores)),
            "n": len(scores),
        }
    
    stats["Mean"] = {
        "mean": float(np.mean([stats[r]["mean"] for r in ["WT", "TC", "ET"]])),
    }
    
    # Print results
    print(f"\nBoundary-band Dice (radius={radius} voxels):")
    print("-" * 40)
    print(f"{'Region':<10} {'Mean':>10} {'Std':>10} {'N':>8}")
    print("-" * 40)
    for region in ["WT", "TC", "ET"]:
        s = stats[region]
        print(f"{region:<10} {s['mean']:>10.4f} {s['std']:>10.4f} {s['n']:>8}")
    print("-" * 40)
    print(f"{'Mean':<10} {stats['Mean']['mean']:>10.4f}")
    
    # Save results
    output_path = os.path.join(output_dir, "table7_boundary_dice.json")
    with open(output_path, "w") as f:
        json.dump({"stats": stats, "config": results["config"]}, f, indent=2)
    print(f"\nResults saved to: {output_path}")
    
    return stats


# ============================================================================
# Table 4: ET-absent Subset Evaluation
# ============================================================================
def eval_table4(predictor: ModelPredictor, val_paths: List[str], 
                config: Dict, output_dir: str) -> Dict:
    """Evaluate Table 4: ET-absent subset evaluation.
    
    Metrics:
    - ET-absent subset (n cases where GT ET voxels = 0):
        - FP volume (mm³)
        - FP components (count)
    - Full validation set:
        - ET presence classification: AUROC, Acc, Sens, Spec
    """
    print("\n" + "=" * 60)
    print("Table 4: ET-absent Subset Evaluation")
    print("=" * 60)
    
    min_size = config.get("et_cc_min_size", 50)
    threshold = config.get("threshold", 0.5)
    
    # Collect results
    et_absent_results = []  # For ET-absent subset
    classification_results = []  # For full validation set
    
    for npz_path in tqdm(val_paths, desc="Evaluating"):
        case = load_case(npz_path)
        if case["label"] is None:
            continue
        
        # Get predictions
        pred_out = predictor.predict(case["image"])
        
        # Get GT regions
        gt_regions = label_to_regions(case["label"])
        gt_et = gt_regions[2]
        gt_et_voxels = int(gt_et.sum())
        
        # Classification labels: y=1 if ET present, 0 otherwise
        y_true = 1 if gt_et_voxels > 0 else 0
        
        # Score for classification: pi_et
        s_score = pred_out["pi_et"]
        
        # Binary ET prediction (after gating + threshold + post-process)
        pred_et_prob = pred_out["probs_gated"][2]
        pred_et_bin = (pred_et_prob > threshold).astype(np.uint8)
        pred_et_bin = remove_small_components(pred_et_bin, min_size, connectivity=3)
        
        # Store classification data
        classification_results.append({
            "case_id": case["case_id"],
            "y_true": y_true,
            "s_score": s_score,
            "y_pred": 1 if s_score > 0.5 else 0,
        })
        
        # For ET-absent cases, compute FP metrics
        if gt_et_voxels == 0:
            fp_voxels = int(pred_et_bin.sum())
            fp_components = count_connected_components(pred_et_bin, connectivity=3)
            
            et_absent_results.append({
                "case_id": case["case_id"],
                "fp_volume_mm3": fp_voxels,  # spacing=1mm
                "fp_components": fp_components,
            })
    
    # -------------------------
    # ET-absent subset metrics
    # -------------------------
    n_et_absent = len(et_absent_results)
    if n_et_absent > 0:
        fp_volumes = [r["fp_volume_mm3"] for r in et_absent_results]
        fp_components = [r["fp_components"] for r in et_absent_results]
        
        et_absent_stats = {
            "n": n_et_absent,
            "fp_volume_mm3": {
                "mean": float(np.mean(fp_volumes)),
                "std": float(np.std(fp_volumes)),
                "min": float(np.min(fp_volumes)),
                "max": float(np.max(fp_volumes)),
            },
            "fp_components": {
                "mean": float(np.mean(fp_components)),
                "std": float(np.std(fp_components)),
                "min": int(np.min(fp_components)),
                "max": int(np.max(fp_components)),
            },
        }
    else:
        et_absent_stats = {"n": 0, "fp_volume_mm3": None, "fp_components": None}
    
    # -------------------------
    # Classification metrics (full validation set)
    # -------------------------
    y_true = np.array([r["y_true"] for r in classification_results])
    s_score = np.array([r["s_score"] for r in classification_results])
    
    # AUROC and optimal threshold using Youden's J statistic
    if len(np.unique(y_true)) > 1:
        auroc = roc_auc_score(y_true, s_score)
        
        # Find optimal threshold using Youden's J = Sens + Spec - 1
        fpr, tpr, thresholds = roc_curve(y_true, s_score)
        j_scores = tpr - fpr  # Youden's J statistic
        best_idx = np.argmax(j_scores)
        optimal_threshold = thresholds[best_idx]
        
        # Use optimal threshold for predictions
        y_pred_optimal = (s_score >= optimal_threshold).astype(int)
    else:
        auroc = float("nan")
        optimal_threshold = 0.5
        y_pred_optimal = (s_score >= 0.5).astype(int)
    
    # Compute metrics at optimal threshold
    tp = int(((y_true == 1) & (y_pred_optimal == 1)).sum())
    tn = int(((y_true == 0) & (y_pred_optimal == 0)).sum())
    fp = int(((y_true == 0) & (y_pred_optimal == 1)).sum())
    fn = int(((y_true == 1) & (y_pred_optimal == 0)).sum())
    
    acc_optimal = (tp + tn) / len(y_true) if len(y_true) > 0 else float("nan")
    sens_optimal = tp / (tp + fn) if (tp + fn) > 0 else float("nan")
    spec_optimal = tn / (tn + fp) if (tn + fp) > 0 else float("nan")
    
    classification_stats = {
        "n": len(classification_results),
        "n_et_present": int(y_true.sum()),
        "n_et_absent": int((1 - y_true).sum()),
        "auroc": float(auroc),
        "optimal_threshold": float(optimal_threshold),
        "accuracy_optimal": float(acc_optimal),
        "sensitivity_optimal": float(sens_optimal),
        "specificity_optimal": float(spec_optimal),
    }
    
    # Print results
    print(f"\nET Presence Classification (n={classification_stats['n']}):")
    print("-" * 50)
    print(f"ET-present: {classification_stats['n_et_present']}, "
          f"ET-absent: {classification_stats['n_et_absent']}")
    print(f"AUROC:              {classification_stats['auroc']:.4f}")
    print(f"Optimal Threshold:  {classification_stats['optimal_threshold']:.4f}")
    print(f"Accuracy:           {classification_stats['accuracy_optimal']:.4f}")
    print(f"Sensitivity:        {classification_stats['sensitivity_optimal']:.4f}")
    print(f"Specificity:        {classification_stats['specificity_optimal']:.4f}")
    
    # Save results
    results = {
        "et_absent_subset": et_absent_stats,
        "et_absent_cases": et_absent_results,
        "classification": classification_stats,
        "config": {"min_size": min_size, "threshold": threshold},
    }
    
    output_path = os.path.join(output_dir, "table4_et_absent.json")
    with open(output_path, "w") as f:
        json.dump(results, f, indent=2)
    print(f"\nResults saved to: {output_path}")
    
    return results


# ============================================================================
# Main
# ============================================================================
def main():
    parser = argparse.ArgumentParser(description="Ablation Evaluation for GliomaSAM3-MoE")
    parser.add_argument("--config", type=str, default="configs/train.yaml",
                        help="Model config path")
    parser.add_argument("--checkpoint", type=str, required=True,
                        help="Model checkpoint path")
    parser.add_argument("--eval", type=str, default="both",
                        choices=["table4", "table7", "both"],
                        help="Which evaluation to run")
    parser.add_argument("--seed", type=int, default=20251225,
                        help="Random seed for data split")
    parser.add_argument("--output_dir", type=str, default="./eval_results",
                        help="Output directory for results")
    parser.add_argument("--device", type=str, default="cuda",
                        help="Device to use")
    parser.add_argument("--data_dir", type=str, default=None,
                        help="Override data directory")
    parser.add_argument("--use_all", action="store_true",
                        help="Use all data instead of validation split only")
    args = parser.parse_args()
    
    # Setup
    os.makedirs(args.output_dir, exist_ok=True)
    
    # Load model config
    model_cfg = load_yaml(args.config)
    
    # Setup evaluation config
    config = DEFAULT_CONFIG.copy()
    config["seed"] = args.seed
    if args.data_dir:
        config["data_dir"] = args.data_dir
    else:
        config["data_dir"] = model_cfg.get("data", {}).get("root_dir", config["data_dir"])
    
    print("=" * 60)
    print("GliomaSAM3-MoE Ablation Evaluation")
    print("=" * 60)
    print(f"Config: {args.config}")
    print(f"Checkpoint: {args.checkpoint}")
    print(f"Data dir: {config['data_dir']}")
    print(f"Seed: {config['seed']}")
    print(f"Evaluation: {args.eval}")
    print(f"Use all data: {args.use_all}")
    
    # Get data paths
    import glob
    all_paths = sorted(glob.glob(os.path.join(config["data_dir"], "*.npz")))
    
    if args.use_all:
        # Use all data for evaluation
        val_paths = all_paths
        print(f"\nUsing all data: {len(val_paths)} cases")
    else:
        # Split data
        print("\nSplitting data...")
        train_paths, val_paths, test_paths = split_npz_paths(
            config["data_dir"],
            train_rate=config["train_rate"],
            val_rate=config["val_rate"],
            test_rate=config["test_rate"],
            seed=config["seed"],
        )
        print(f"Train: {len(train_paths)}, Val: {len(val_paths)}, Test: {len(test_paths)}")
    
    # Initialize predictor
    print("\nLoading model...")
    predictor = ModelPredictor(args.config, args.checkpoint, args.device)
    
    # Run evaluations
    results = {}
    
    if args.eval in ["table7", "both"]:
        results["table7"] = eval_table7(predictor, val_paths, config, args.output_dir)
    
    if args.eval in ["table4", "both"]:
        results["table4"] = eval_table4(predictor, val_paths, config, args.output_dir)
    
    print("\n" + "=" * 60)
    print("Evaluation Complete!")
    print("=" * 60)
    
    return results


if __name__ == "__main__":
    main()