Benchmarking CNNs, Vision Transformers, and Foundation Models for Explainable Brain Tumor MRI Classification: Calibration, Decision Analysis, and External Validation
Organizations: Department of Computer Science, Faculty of Computing and Information Systems, Ghana Communication Technology University, Accra, Ghana
Abstract
Deep learning models for brain tumor MRI classification are usually compared by accuracy alone, on public datasets whose independence is rarely checked. We benchmarked six architectures: two convolutional networks (EfficientNetV2-M and ConvNeXt-V2-Tiny), two vision transformers (Swin-V2-S and ViT-L/16), and two foundation models (DINOv2-L and MedSigLIP-448), for four-class classification of 7,216 public MRI images. We assessed accuracy, calibration, explainability, ensembles, and decision curves, with every fitted quantity estimated on validation data only. Pixel-verified duplicate screening showed that 36% of the test images had a near-duplicate in the training data. Test accuracy ranged from 94.5% to 95.6%, with no pairwise difference after Holm correction, and fell to 90.3% to 91.7% on the 773 test images free of duplicates and shared patients. Ensembles did not outperform the best single model. Attribution maps passed a model-randomization test but localized expert-annotated tumors poorly. On 1,675 external BDNeuro-MRI images that did not duplicate the training data, accuracy ranged from 77.6% (MedSigLIP-448) to 95.0% (ViT-L/16), and 12 of 15 pairwise differences were significant. Internal accuracy on this public collection therefore did not distinguish the architectures, whereas the duplicate-filtered external test did.
Figures & tables
| Study | Architecture(s) | Task | Data as described by the authors | Reported result |
|---|---|---|---|---|
| Lall et al. [ 32 ] | ResNet-50, VGG-19, custom CNN | 3-class classification | Brain MRI (glioma, meningioma, pituitary) | Accuracy 97.91% (ResNet-50) |
| Younis et al. [ 31 ] | CNN, VGG-16, ensemble | Binary detection | 253 images (155 with tumor) | Accuracy 98.14% (ensemble); precision 98.41% |
| Ali et al. [ 33 ] | ResNet50, classic CNN | Multi-class classification | Brain MRI | Accuracy 99.88% (ResNet50) |
| Abdusalomov et al. [ 30 ] | YOLOv7 with CBAM, SPPF+, BiFPN | 3-class detection | Augmented to 51,448 images | Detection accuracy 99.5% |
| Atmakuri et al. [ 34 ] | ResNet-50, CNN, ANN | Tumor phase prediction | Kaggle dataset | Accuracy about 95% |
| Shahzad et al. [ 35 ] | Tumor-Swin-Transformer | 4-class classification | Multi-source dataset | Accuracy 98% |
| Class | Kaggle Training folder | Training (this study) | Validation (this study) | Kaggle Testing folder = test set | Total |
|---|---|---|---|---|---|
| Glioma | 1,400 | 1,129 | 271 | 400 | 1,800 |
| Meningioma | 1,400 | 1,151 | 249 | 400 | 1,800 |
| No tumor | 1,416 | 1,124 | 292 | 400 | 1,816 |
| Pituitary | 1,400 | 1,119 | 281 | 400 | 1,800 |
| Total | 5,616 | 4,523 | 1,093 | 1,600 | 7,216 |
| Partition | Class | Figshare | SARTAJ | Br35H | Unmatched | Total |
|---|---|---|---|---|---|---|
| Training | Glioma | 371 | 67 | 14 | 677 | 1,129 |
| Meningioma | 602 | 144 | 52 | 353 | 1,151 | |
| No tumor | 0 | 150 | 974 | 0 | 1,124 | |
| Pituitary | 533 | 151 | 8 | 427 | 1,119 | |
| Validation | Glioma | 82 | 8 | 4 | 177 | 271 |
| Meningioma | 102 | 30 | 8 | 109 | 249 |
| Quantity | Value |
|---|---|
| Images in the Kaggle composite | 7,216 (Training folder 5,616; Testing folder 1,600) |
| Exact-duplicate groups (images) | 288 (626) |
| Candidate near-duplicate pairs (pHash Hamming distance 6) | 8,922 (median NCC 0.978; 5th percentile 0.705) |
| Verified near-duplicate pairs (NCC 0.95) | 5,786 |
| of which cross the Training/Testing folders | 2,077 |
| of which have conflicting class labels | 10 |
| Model | Family / pretraining | Params (M) | Input (px) | Normalisation | Optimiser, LR, WD | Batch | Best / stop epoch |
|---|---|---|---|---|---|---|---|
| EfficientNetV2-M | CNN, ImageNet-1K | 52.86 | 224 | ImageNet | Adam, 1e-4, 1e-5 | 32 | 17 / 18 |
| ConvNeXt-V2-Tiny | CNN, FCMAE + IN-22K/1K | 27.87 | 224 | ImageNet | AdamW, 5e-5, 0.01 | 32 | 5 / 11 |
| Swin-V2-S | Hierarchical ViT, IN-1K | 48.96 | 256 | ImageNet | AdamW, 5e-5, 0.01 | 32 | 6 / 14 |
| ViT-L/16 | ViT, IN-21K | 303.31 | 224 | ImageNet | AdamW, 3e-5, 0.01 | 32 | 8 / 18 |
| DINOv2-L | ViT, self-supervised (LVD-142M) | 304.38 | 224 | ImageNet | AdamW, 2e-5, 0.01 | 32 | 6 / 11 |
| MedSigLIP-448 | ViT, medical image-text (SigLIP) | 428.57 | 448 | mean = SD = 0.5 | AdamW, 2e-5, 0.01 | 8 | 22 / 27 |
| Model | Validation accuracy | Test accuracy [95% CI] | Macro-F1 [95% CI] | Macro AUC [95% CI] | MCC [95% CI] | Clean-test accuracy (n = 773) |
|---|---|---|---|---|---|---|
| EfficientNetV2-M | 0.9909 | 0.9556 [0.9450, 0.9656] | 0.9548 [0.9441, 0.9648] | 0.9795 [0.9734, 0.9850] | 0.9422 [0.9289, 0.9550] | 0.9133 |
| ConvNeXt-V2-Tiny | 0.9835 | 0.9500 [0.9394, 0.9600] | 0.9490 [0.9382, 0.9591] | 0.9862 [0.9806, 0.9910] | 0.9351 [0.9215, 0.9478] | 0.9082 |
| Swin-V2-S | 0.9863 | 0.9450 [0.9331, 0.9556] | 0.9441 [0.9323, 0.9547] | 0.9850 [0.9798, 0.9897] | 0.9284 [0.9136, 0.9419] | 0.9030 |
| ConvNeXt-V2-Tiny (448 px) | 0.9872 | 0.9587 [0.9481, 0.9675] | 0.9580 [0.9474, 0.9672] | 0.9810 [0.9747, 0.9870] | 0.9462 [0.9328, 0.9576] | 0.9237 |
| ConvNeXt-V2-Tiny (448 px, matched) | 0.9826 | 0.9419 [0.9306, 0.9525] | 0.9402 [0.9286, 0.9513] | 0.9735 [0.9665, 0.9804] | 0.9250 [0.9111, 0.9385] | 0.8939 |
| ViT-L/16 | 0.9835 | 0.9525 [0.9419, 0.9625] | 0.9518 [0.9411, 0.9616] | 0.9853 [0.9802, 0.9900] | 0.9381 [0.9245, 0.9508] | 0.9172 |
| Model | Glioma | Meningioma | No tumor | Pituitary |
|---|---|---|---|---|
| EfficientNetV2-M | 0.914 / 0.843 / 0.939 | 0.944 / 0.985 / 0.982 | 0.972 / 1.000 / 0.998 | 0.989 / 0.995 / 0.999 |
| ConvNeXt-V2-Tiny | 0.899 / 0.823 / 0.964 | 0.932 / 0.988 / 0.984 | 0.977 / 0.998 / 0.999 | 0.989 / 0.993 / 0.998 |
| Swin-V2-S | 0.900 / 0.825 / 0.951 | 0.947 / 0.965 / 0.996 | 0.941 / 1.000 / 0.994 | 0.988 / 0.990 / 0.999 |
| ViT-L/16 | 0.911 / 0.840 / 0.958 | 0.929 / 0.978 / 0.985 | 0.985 / 0.993 / 0.999 | 0.983 / 1.000 / 1.000 |
| DINOv2-L | 0.895 / 0.823 / 0.937 | 0.925 / 0.993 / 0.983 | 0.982 / 0.995 / 1.000 | 0.985 / 0.980 / 1.000 |
| MedSigLIP-448 | 0.906 / 0.828 / 0.949 | 0.938 / 0.988 / 0.993 | 0.972 / 1.000 / 0.999 | 0.995 / 1.000 / 1.000 |
| Model | Figshare (n = 305) | SARTAJ (n = 172) | Br35H (n = 487) | Unmatched (n = 636) | Clean subset (n = 773) |
|---|---|---|---|---|---|
| EfficientNetV2-M | 0.9967 | 0.8547 | 0.9220 | 0.9890 | 0.9133 |
| ConvNeXt-V2-Tiny | 0.9869 | 0.8140 | 0.9261 | 0.9874 | 0.9082 |
| Swin-V2-S | 0.9869 | 0.8081 | 0.9179 | 0.9827 | 0.9030 |
| ViT-L/16 | 0.9934 | 0.8023 | 0.9261 | 0.9937 | 0.9172 |
| DINOv2-L | 0.9836 | 0.8023 | 0.9281 | 0.9843 | 0.9082 |
| MedSigLIP-448 | 0.9967 | 0.8256 | 0.9220 | 0.9921 | 0.9120 |
| Comparison | Accuracy difference (pp) | 95% paired-bootstrap CI (pp) | Discordant b / c | Exact p | Holm-adjusted p |
|---|---|---|---|---|---|
| EfficientNetV2-M vs ConvNeXt-V2-Tiny | +0.56 | [-0.06, +1.25] | 19 / 10 | 0.1360 | 1.0000 |
| EfficientNetV2-M vs Swin-V2-S | +1.06 | [+0.31, +1.81] | 26 / 9 | 0.0060 | 0.0898 |
| EfficientNetV2-M vs ViT-L/16 | +0.31 | [-0.31, +0.94] | 16 / 11 | 0.4421 | 1.0000 |
| EfficientNetV2-M vs DINOv2-L | +0.81 | [+0.06, +1.56] | 24 / 11 | 0.0410 | 0.5325 |
| EfficientNetV2-M vs MedSigLIP-448 | +0.19 | [-0.37, +0.75] | 12 / 9 | 0.6636 | 1.0000 |
| ConvNeXt-V2-Tiny vs Swin-V2-S | +0.50 | [-0.19, +1.25] | 21 / 13 | 0.2295 | 1.0000 |
| Model | Temperature T | ECE (15 bins), uncalibrated | ECE (15 bins), temperature-scaled | Equal-mass ECE, scaled | Class-wise ECE, scaled | Brier (uncal. scaled) | NLL (uncal. scaled) |
|---|---|---|---|---|---|---|---|
| EfficientNetV2-M | 0.578 | 0.0403 [0.0312, 0.0499] | 0.0372 [0.0289, 0.0487] | 0.0347 [0.0261, 0.0460] | 0.0215 [0.0173, 0.0269] | 0.0848 0.0865 | 0.2278 0.2624 |
| ConvNeXt-V2-Tiny | 0.626 | 0.0579 [0.0498, 0.0676] | 0.0349 [0.0265, 0.0456] | 0.0370 [0.0305, 0.0492] | 0.0221 [0.0177, 0.0276] | 0.0888 0.0915 | 0.2331 0.2442 |
| Swin-V2-S | 0.622 | 0.0526 [0.0422, 0.0622] | 0.0402 [0.0309, 0.0517] | 0.0414 [0.0346, 0.0546] | 0.0243 [0.0200, 0.0302] | 0.0982 0.1020 | 0.2525 0.2769 |
| ConvNeXt-V2-Tiny (448 px) | 0.594 | 0.0589 [0.0508, 0.0684] | 0.0302 [0.0222, 0.0407] | 0.0326 [0.0267, 0.0449] | 0.0190 [0.0149, 0.0247] | 0.0775 0.0775 | 0.2194 0.2316 |
| ConvNeXt-V2-Tiny (448 px, matched) | 0.681 | 0.0345 [0.0235, 0.0451] | 0.0397 [0.0308, 0.0507] | 0.0434 [0.0368, 0.0564] | 0.0274 [0.0228, 0.0330] | 0.1088 0.1111 | 0.2806 0.3035 |
| ViT-L/16 | 0.640 | 0.0621 [0.0547, 0.0723] | 0.0307 [0.0221, 0.0410] | 0.0369 [0.0304, 0.0488] | 0.0204 [0.0164, 0.0258] | 0.0860 0.0877 | 0.2321 0.2343 |
| Method | Parameters fitted on | Accuracy [95% CI] | Macro-F1 | Macro AUC | Log-loss | vs best single, b / c | Exact p | Identical to best single |
|---|---|---|---|---|---|---|---|---|
| Simple average | none | 0.9544 [0.9437, 0.9644] | 0.9534 | 0.9929 | 0.2072 | 8 / 10 | 0.8145 | 0.984 |
| Weighted average | validation | 0.9550 [0.9450, 0.9644] | 0.9541 | 0.9892 | 0.2344 | 9 / 10 | 1.0000 | 0.978 |
| Majority vote | none | 0.9544 [0.9437, 0.9644] | 0.9534 | 0.9929 | 0.4219 | 8 / 10 | 0.8145 | 0.984 |
| Stacking | validation | 0.9575 [0.9475, 0.9669] | 0.9569 | 0.9893 | 0.2364 | 9 / 6 | 0.6072 | 0.984 |
| Best single on validation (MedSigLIP-448) | validation | 0.9538 [0.9431, 0.9637] | 0.9527 | 0.9853 | 0.2702 | 9 / 12 | 0.6636 | 0.976 |
| Weighted average | test, 5-fold cross-fitted (sensitivity) | 0.9525 [0.9419, 0.9625] | 0.9516 | 0.9916 | 0.2077 | 7 / 12 | 0.3593 | – |
| Model | Argmax (no threshold) | Scaled argmax, argmax(p_k / t_k) | One-flag rule |
|---|---|---|---|
| EfficientNetV2-M | 0.9556 | 0.9519 | 0.9512 |
| ConvNeXt-V2-Tiny | 0.9500 | 0.9450 | 0.9488 |
| Swin-V2-S | 0.9450 | 0.9438 | 0.9450 |
| ViT-L/16 | 0.9525 | 0.9500 | 0.9519 |
| DINOv2-L | 0.9475 | 0.9438 | 0.9469 |
| MedSigLIP-448 | 0.9538 | 0.9575 | 0.9569 |
| Model | Glioma | Meningioma | No tumor | Pituitary |
|---|---|---|---|---|
| EfficientNetV2-M | 0.792 (0.825 / 1.000) | 0.117 (0.990 / 0.959) | 0.367 (1.000 / 0.978) | 0.706 (0.995 / 0.995) |
| ConvNeXt-V2-Tiny | 0.750 (0.815 / 0.998) | 0.026 (0.990 / 0.938) | 0.123 (0.998 / 0.982) | 0.024 (0.998 / 0.992) |
| Swin-V2-S | 0.084 (0.825 / 0.996) | 0.006 (0.990 / 0.948) | 0.265 (1.000 / 0.956) | 0.931 (0.978 / 0.997) |
| ViT-L/16 | 0.827 (0.838 / 0.998) | 0.024 (0.990 / 0.943) | 0.007 (0.998 / 0.965) | 0.799 (1.000 / 0.990) |
| DINOv2-L | 0.075 (0.833 / 0.991) | 0.102 (0.993 / 0.942) | 0.037 (0.995 / 0.983) | 0.716 (0.973 / 0.997) |
| MedSigLIP-448 | 0.021 (0.843 / 1.000) | 0.844 (0.988 / 0.962) | 0.965 (1.000 / 0.985) | 0.993 (0.998 / 0.999) |
| Comparison (only the stated factor differs) | Accuracy A | Accuracy B | Difference A B (pp) [95% CI] | Discordant b / c | Exact p |
|---|---|---|---|---|---|
| MedSigLIP-448 vs MedSigLIP (224 px) | 0.9538 | 0.9531 | +0.06 [-0.56, +0.63] | 13 / 12 | 1.0000 |
| ConvNeXt-V2-Tiny (448 px) vs ConvNeXt-V2-Tiny | 0.9588 | 0.9500 | +0.88 [+0.37, +1.44] | 17 / 3 | 0.0026 |
| MedSigLIP (224 px) vs ConvNeXt-V2-Tiny | 0.9531 | 0.9500 | +0.31 [-0.37, +1.00] | 18 / 13 | 0.4731 |
| MedSigLIP-448 vs ConvNeXt-V2-Tiny (448 px) | 0.9538 | 0.9588 | -0.50 [-1.06, +0.00] | 6 / 14 | 0.1153 |
| MedSigLIP-448 vs ConvNeXt-V2-Tiny (448 px, matched) | 0.9538 | 0.9419 | +1.19 [+0.56, +1.81] | 23 / 4 | 0.0003 |
| ConvNeXt-V2-Tiny (448 px, matched) vs ConvNeXt-V2-Tiny (448 px) | 0.9419 | 0.9588 | -1.69 [-2.50, -1.00] | 6 / 33 | <0.0001 |
| Fine-tuned model | Full image | Brain only | Change (pp) | p (brain vs full) | Background only | Source prediction (Figshare vs SARTAJ) † |
|---|---|---|---|---|---|---|
| EfficientNetV2-M | 0.9556 | 0.9444 | -1.12 | 0.0003 | 0.2581 | 0.8302 |
| ConvNeXt-V2-Tiny | 0.9500 | 0.9406 | -0.94 | 0.0015 | 0.2769 | 0.8477 |
| Swin-V2-S | 0.9450 | 0.9319 | -1.31 | 0.0000 | 0.3988 | 0.8372 |
| ConvNeXt-V2-Tiny (448 px) | 0.9588 | 0.9456 | -1.32 | 0.0000 | 0.4031 | – |
| ConvNeXt-V2-Tiny (448 px, matched) | 0.9419 | 0.9338 | -0.81 | 0.0072 | 0.3188 | – |
| ViT-L/16 | 0.9525 | 0.9462 | -0.63 | 0.0063 | 0.2744 | 0.8233 |
| Frozen pretrained backbone + logistic regression | Full image | Brain only | Background only |
|---|---|---|---|
| ConvNeXt-V2-Tiny | 0.9150 | 0.9081 | 0.8406 |
| DINOv2-L | 0.9300 | 0.9338 | 0.8850 |
| MedSigLIP-448 | 0.9381 | 0.9438 | 0.8531 |
| Model | Method | n | Pointing-game hit rate | Energy in mask | Chance (mask area) | Energy / chance | IoU (top-20% map) |
|---|---|---|---|---|---|---|---|
| EfficientNetV2-M | Grad-CAM | 165 | 0.139 | 0.085 | 0.016 | 5.25 | 0.076 |
| Grad-CAM++ | 165 | 0.145 | 0.092 | 0.016 | 5.63 | 0.076 | |
| EigenCAM | 165 | 0.139 | 0.094 | 0.016 | 5.77 | 0.076 | |
| Integrated Gradients | 40 | 0.000 | 0.045 | 0.021 | 2.12 | 0.055 | |
| LIME | 20 | 0.050 | 0.048 | 0.020 | 2.37 | 0.055 | |
| ConvNeXt-V2-Tiny | Grad-CAM | 165 | 0.255 | 0.095 | 0.016 | 5.82 | 0.073 |
| Model | n (occlusion) | Occlusion pointing | Occlusion energy in mask | Occlusion energy / chance | Grad-CAM energy in brain (%) | Grad-CAM energy in outer 10% border (%) | Grad-CAM energy in tumor mask (%) |
| ConvNeXt-V2-Tiny | 40 | 0.050 | 0.038 | 1.77 | 91.6 | 3.9 | 9.6 |
| DINOv2-L | 40 | 0.275 | 0.085 | 3.97 | 40.4 | 41.5 | 1.3 |
| EfficientNetV2-M | 40 | 0.150 | 0.065 | 3.03 | 98.8 | 0.5 | 11.0 |
| Ensemble (weighted average) | 40 | 0.075 | 0.060 | 2.80 | – | – | – |
| MedSigLIP-448 | 40 | 0.025 | 0.040 | 1.88 | 49.8 | 35.2 | 2.6 |
| Swin-V2-S | 40 | 0.100 | 0.027 | 1.25 | 96.6 | 0.8 | 6.4 |
| Model | Head randomised | + last stage / block | + 2 stages / 4 blocks | + 3 stages / 12 blocks | Stability: Gaussian noise | Stability: horizontal flip |
|---|---|---|---|---|---|---|
| EfficientNetV2-M | -0.696 | -0.014 | -0.207 | -0.192 | 0.956 | 0.853 |
| ConvNeXt-V2-Tiny | 0.546 | 0.174 | -0.049 | -0.123 | 0.918 | 0.921 |
| Swin-V2-S | 0.434 | -0.116 | -0.172 | -0.125 | 0.978 | 0.966 |
| ViT-L/16 | -0.126 | 0.090 | 0.133 | 0.034 | 0.847 | 0.720 |
| DINOv2-L | -0.349 | -0.237 | -0.434 | -0.307 | 0.978 | 0.768 |
| MedSigLIP-448 | 0.043 | 0.120 | 0.151 | 0.189 | 0.592 | 0.689 |
| Model | Threshold p_t | Net benefit [95% CI] (maximum 0.100) | Unnecessary prioritisations avoided per 100 vs prioritise-all | Sensitivity | Specificity | Missed tumors per 1,000 | Unnecessary prioritisations per 1,000 |
|---|---|---|---|---|---|---|---|
| EfficientNetV2-M | 0.05 | 0.0979 [0.0970, 0.0987] | 86.0 | 0.979 | 1.000 | 2.1 | 0.0 |
| 0.10 | 0.0977 [0.0968, 0.0985] | 88.0 | 0.978 | 1.000 | 2.2 | 0.0 | |
| 0.20 | 0.0977 [0.0968, 0.0985] | 89.1 | 0.978 | 1.000 | 2.2 | 0.0 | |
| ConvNeXt-V2-Tiny | 0.05 | 0.0984 [0.0976, 0.0991] | 86.9 | 0.985 | 0.998 | 1.5 | 2.2 |
| 0.10 | 0.0982 [0.0972, 0.0990] | 88.3 | 0.984 | 0.998 | 1.6 | 2.2 | |
| 0.20 | 0.0976 [0.0961, 0.0987] | 89.0 | 0.982 | 0.998 | 1.8 | 2.2 |
| Model | Glioma | Meningioma | No tumor | Pituitary |
|---|---|---|---|---|
| EfficientNetV2-M | 12,761 | 1,856 | 40,135 | 33,915 |
| ConvNeXt-V2-Tiny | 1,849 | 1,710 | 26,201 | 26,334 |
| Swin-V2-S | 1,881 | 1,113 | 18,248 | 19,701 |
| ViT-L/16 | 3,145 | 979 | 17,512 | 65,544 |
| DINOv2-L | 922 | 2,471 | 18,170 | 14,651 |
| MedSigLIP-448 | 11,452 | 1,938 | 40,135 | 212,977 |
| Model | AUPRC glioma | AUPRC meningioma | AUPRC no tumor | AUPRC pituitary | Macro AUPRC | Accuracy at 80% coverage | at 90% | at 95% | at 100% |
|---|---|---|---|---|---|---|---|---|---|
| EfficientNetV2-M | 0.930 | 0.886 | 0.992 | 0.996 | 0.951 | 0.9734 | 0.9701 | 0.9671 | 0.9556 |
| ConvNeXt-V2-Tiny | 0.958 | 0.912 | 0.995 | 0.985 | 0.963 | 0.9828 | 0.9799 | 0.9776 | 0.9500 |
| Swin-V2-S | 0.943 | 0.986 | 0.956 | 0.998 | 0.971 | 0.9859 | 0.9764 | 0.9664 | 0.9450 |
| ViT-L/16 | 0.947 | 0.896 | 0.994 | 0.999 | 0.959 | 0.9820 | 0.9806 | 0.9757 | 0.9525 |
| DINOv2-L | 0.931 | 0.883 | 0.999 | 0.999 | 0.953 | 0.9789 | 0.9750 | 0.9697 | 0.9475 |
| MedSigLIP-448 | 0.951 | 0.975 | 0.993 | 1.000 | 0.980 | 0.9914 | 0.9854 | 0.9796 | 0.9538 |
| Model / ensemble | Accuracy [95% CI] | Balanced accuracy [95% CI] | Macro-F1 [95% CI] | Macro AUC [95% CI] | MCC | ECE | Internal test accuracy | Change (pp) | Strict set accuracy (n = 1,485) |
|---|---|---|---|---|---|---|---|---|---|
| EfficientNetV2-M | 0.9296 [0.9164, 0.9415] | 0.9355 [0.9227, 0.9471] | 0.9150 [0.8994, 0.9295] | 0.9865 [0.9822, 0.9904] | 0.8980 | 0.0545 | 0.9556 | -2.61 | 0.9239 |
| ConvNeXt-V2-Tiny | 0.9313 [0.9194, 0.9439] | 0.9308 [0.9173, 0.9444] | 0.9195 [0.9044, 0.9340] | 0.9917 [0.9886, 0.9946] | 0.8988 | 0.0472 | 0.9500 | -1.87 | 0.9259 |
| Swin-V2-S | 0.8603 [0.8430, 0.8764] | 0.8612 [0.8440, 0.8776] | 0.8260 [0.8044, 0.8461] | 0.9867 [0.9831, 0.9902] | 0.8086 | 0.1001 | 0.9450 | -8.47 | 0.8572 |
| ViT-L/16 | 0.9499 [0.9397, 0.9606] | 0.9521 [0.9406, 0.9633] | 0.9443 [0.9325, 0.9560] | 0.9917 [0.9887, 0.9945] | 0.9251 | 0.0306 | 0.9525 | -0.26 | 0.9461 |
| DINOv2-L | 0.9272 [0.9146, 0.9397] | 0.9247 [0.9098, 0.9387] | 0.9200 [0.9057, 0.9335] | 0.9878 [0.9841, 0.9913] | 0.8918 | 0.0556 | 0.9475 | -2.03 | 0.9239 |
| MedSigLIP-448 | 0.7761 [0.7564, 0.7952] | 0.8340 [0.8175, 0.8497] | 0.7767 [0.7563, 0.7960] | 0.9809 [0.9766, 0.9850] | 0.7239 | 0.2054 | 0.9538 | -17.76 | 0.7690 |
| Model / ensemble | Glioma (n = 862) | Meningioma (n = 306) | No tumor (n = 222) | Pituitary (n = 285) |
|---|---|---|---|---|
| EfficientNetV2-M | 0.922 / 0.999 | 0.994 / 0.953 | 0.991 / 0.964 | 0.835 / 1.000 |
| ConvNeXt-V2-Tiny | 0.934 / 0.985 | 0.974 / 0.947 | 0.960 / 0.979 | 0.856 / 1.000 |
| Swin-V2-S | 0.870 / 1.000 | 0.964 / 0.971 | 1.000 / 0.868 | 0.611 / 0.999 |
| ViT-L/16 | 0.942 / 0.984 | 0.977 / 0.964 | 0.928 / 0.992 | 0.961 / 0.993 |
| DINOv2-L | 0.928 / 0.967 | 0.984 / 0.942 | 0.910 / 0.992 | 0.877 / 0.997 |
| MedSigLIP-448 | 0.679 / 1.000 | 0.984 / 0.963 | 1.000 / 0.777 | 0.674 / 1.000 |
| Model | Parameters (M) | Input (px) | Inference latency, batch 1 (ms, median) | Training time per epoch (s) | Peak training GPU memory (GB) |
|---|---|---|---|---|---|
| EfficientNetV2-M | 52.86 | 224 | 37.4 | 24.6 | 5.1 |
| ConvNeXt-V2-Tiny | 27.87 | 224 | 12.0 | 10.9 | 4.2 |
| Swin-V2-S | 48.96 | 256 | 39.8 | 23.9 | 6.9 |
| ConvNeXt-V2-Tiny (448 px) | 27.87 | 448 | 12.3 | 29.8 | 15.1 |
| ConvNeXt-V2-Tiny (448 px, matched) | 27.87 | 448 | 12.2 | 38.0 | 5.2 |
| ViT-L/16 | 303.31 | 224 | 16.8 | 16.0 | 11.3 |
Brain Tumor Classification
Brain tumor classification uses medical images and molecular data to identify tumor types, grades, and subtypes for diagnosis and treatment planning.
See all 55 papers