Update README.md
Browse files
README.md
CHANGED
|
@@ -71,7 +71,27 @@ This model is intended for research purposes in the field of neuropathology.
|
|
| 71 |
* Epochs/Iterations: 5000 Iterations
|
| 72 |
* Optimizer: AdamW
|
| 73 |
* Weight decay: 0.04-0.4
|
| 74 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 75 |
## Evaluation
|
| 76 |
|
| 77 |
* **Task(s):** Classification, KNN, Clustering, Robustness
|
|
|
|
| 71 |
* Epochs/Iterations: 5000 Iterations
|
| 72 |
* Optimizer: AdamW
|
| 73 |
* Weight decay: 0.04-0.4
|
| 74 |
+
### Cross-Magnification Embedding Visualization
|
| 75 |
+
|
| 76 |
+
The following PCA visualizations illustrate how embeddings extracted at 10× and 40× magnifications relate to each other across different models. Gray lines connect spatially aligned parent–child tile pairs, enabling a direct comparison of cross-magnification consistency. The DINOv2-Giant baseline shown here is fine-tuned on the same neuropathology dataset using the standard DINO self-supervised training strategy without magnification-aware alignment. While baseline models exhibit clear magnification-dependent separation, MAD-NP produces overlapping clusters where tissue identity is preserved across resolutions, indicating a unified and magnification-stable embedding space.
|
| 77 |
+
<table>
|
| 78 |
+
<tr>
|
| 79 |
+
<td align="center"><b>(a) MAD-NP</b></td>
|
| 80 |
+
<td align="center"><b>(b) Virchow2</b></td>
|
| 81 |
+
<td align="center"><b>(c) DINOv2 Giant Finetuned</b></td>
|
| 82 |
+
</tr>
|
| 83 |
+
<tr>
|
| 84 |
+
<td>
|
| 85 |
+
<img src="MAD-NP_pca_cross_mag.png" width="250">
|
| 86 |
+
</td>
|
| 87 |
+
<td>
|
| 88 |
+
<img src="Virchow2_pca_cross_mag.png" width="250">
|
| 89 |
+
</td>
|
| 90 |
+
<td>
|
| 91 |
+
<img src="DINOv2-Giant_pca_cross_mag.png" width="250">
|
| 92 |
+
</td>
|
| 93 |
+
</tr>
|
| 94 |
+
</table>
|
| 95 |
## Evaluation
|
| 96 |
|
| 97 |
* **Task(s):** Classification, KNN, Clustering, Robustness
|