# Table 3 — Main comparison

| Group       | Method               | Input   | accuracy   | balanced_accuracy   | macro_precision   | macro_recall   | macro_f1   | macro_auc   | macro_f1_CI   |
|:------------|:---------------------|:--------|:-----------|:--------------------|:------------------|:---------------|:-----------|:------------|:--------------|
| Metadata    | XGBoost              | Tabular | 52.4±6.6   | 47.3±6.6            | 49.0±6.1          | 47.3±6.6       | 47.4±6.5   | 66.7±7.4    | [43.5, 52.3]  |
| Metadata    | MLP                  | Tabular | 41.6±10.4  | 44.2±9.2            | 41.6±9.5          | 44.2±9.2       | 38.9±9.7   | 59.1±9.1    | [36.2, 44.4]  |
| CNN         | ResNet50             | 2D      | 44.6±7.4   | 39.9±6.8            | 39.1±8.4          | 39.9±6.8       | 36.8±7.4   | 57.6±6.9    | [35.3, 43.2]  |
| CNN         | DenseNet121          | 2.5D    | 51.4±6.9   | 46.8±7.6            | 47.3±10.8         | 46.8±7.6       | 44.9±7.7   | 67.9±6.6    | [41.3, 49.6]  |
| CNN         | 3D ResNet18          | 3D      | 49.2±4.9   | 41.3±6.2            | 34.6±11.4         | 41.3±6.2       | 35.4±6.8   | 67.2±5.6    | [36.4, 44.7]  |
| Transformer | ViT-B/16             | 2.5D    | 53.7±5.9   | 49.0±7.8            | 49.6±10.5         | 49.0±7.8       | 46.2±8.5   | 69.5±4.9    | [44.2, 52.7]  |
| Transformer | 3D Swin-T            | 3D      | 48.1±6.8   | 44.2±6.7            | 43.7±11.9         | 44.2±6.7       | 40.3±8.5   | 67.9±4.3    | [39.2, 47.7]  |
| Hybrid      | 3D HCCT              | 3D      | 51.9±6.2   | 45.0±7.5            | 41.0±9.6          | 45.0±7.5       | 40.7±7.8   | 64.4±5.5    | [40.7, 49.5]  |
| Hybrid      | CNN-VSwinFormer-lite | 3D      | 43.8±6.8   | 40.0±6.5            | 37.0±10.1         | 40.0±6.5       | 35.4±7.8   | 63.2±5.3    | [34.8, 43.1]  |
| Multimodal  | DenseNet + concat    | MRI+tab | 54.6±7.0   | 52.5±6.6            | 52.1±9.3          | 52.5±6.6       | 49.7±6.2   | 71.0±4.7    | [45.9, 54.3]  |
| Proposed    | TriFuse-AD           | MRI+tab | 53.2±7.3   | 50.8±6.5            | 50.7±7.5          | 50.8±6.5       | 48.8±6.6   | 69.2±6.4    | [45.8, 54.0]  |


# Table 3b — Per-class F1 / Recall

| Method               | f1_CN     | f1_VMD    | f1_AD     | recall_CN   | recall_VMD   | recall_AD   |
|:---------------------|:----------|:----------|:----------|:------------|:-------------|:------------|
| XGBoost              | 63.3±7.6  | 41.3±7.8  | 37.7±13.2 | 66.2±9.6    | 39.9±8.9     | 35.8±13.9   |
| MLP                  | 46.4±17.0 | 29.2±16.4 | 41.2±14.0 | 45.8±22.4   | 29.1±21.8    | 57.8±18.7   |
| ResNet50             | 56.1±13.2 | 33.9±16.1 | 20.5±12.7 | 56.4±18.5   | 35.9±21.3    | 27.3±20.6   |
| DenseNet121          | 66.7±5.4  | 31.6±14.4 | 36.5±11.0 | 71.1±9.2    | 28.4±14.5    | 41.0±17.5   |
| 3D ResNet18          | 61.9±10.6 | 20.4±20.4 | 24.0±21.1 | 74.9±22.3   | 22.3±25.3    | 26.6±25.6   |
| ViT-B/16             | 64.3±9.3  | 41.9±14.6 | 32.3±15.3 | 66.5±19.6   | 44.2±19.8    | 36.3±25.4   |
| 3D Swin-T            | 58.3±12.2 | 34.7±18.4 | 28.0±16.3 | 58.7±20.7   | 40.9±28.4    | 33.0±23.9   |
| 3D HCCT              | 64.6±9.8  | 29.8±21.3 | 27.9±18.3 | 71.2±19.4   | 33.1±28.6    | 30.6±24.3   |
| CNN-VSwinFormer-lite | 54.1±18.8 | 27.9±14.8 | 24.2±16.7 | 58.1±26.2   | 28.2±22.8    | 33.8±27.6   |
| DenseNet + concat    | 67.3±7.7  | 42.8±7.8  | 39.0±12.2 | 67.7±14.6   | 38.7±8.4     | 51.2±24.3   |
| TriFuse-AD           | 62.3±10.1 | 47.4±12.7 | 36.7±12.0 | 57.2±13.1   | 51.1±18.0    | 44.0±21.0   |


# Table 4 — Ablation

| Variant                        | macro_f1   | balanced_accuracy   | f1_AD     |
|:-------------------------------|:-----------|:--------------------|:----------|
| A1: axial-only                 | 47.7±7.1   | 49.7±7.0            | 33.8±13.5 |
| A2: mean-pool (no Transformer) | 46.5±8.3   | 49.7±7.6            | 36.7±14.2 |
| A3: no metadata                | 39.0±10.3  | 43.1±9.1            | 23.4±19.2 |
| A4: concat (no gate)           | 48.7±5.3   | 51.4±6.6            | 37.3±15.8 |
| A5: weighted-CE                | 50.0±5.4   | 51.8±5.2            | 41.3±13.4 |
| Full TriFuse-AD                | 48.8±6.6   | 50.8±6.5            | 36.7±12.0 |


# Significance
TriFuse-AD 48.8 vs densenet_latefusion 49.7 (Δ=-0.9, permutation p=0.5544)