⚠ For research and demonstration purposes only. Not for clinical use.
Capabilities
What Galacticos Can Do
Image Classification
Identify tissue type across 12 dermatological classes with per-class probability scoring and confidence levels.
Tissue Segmentation
4 visualization modes — binary mask, overlay, contour lines, and probability heatmap — powered by nnU-Net.
AI Image Validation
Gemini-powered pre-screening ensures only valid medical images are analyzed, rejecting non-medical uploads.
Drag & Drop Upload
Upload PNG, JPG, or BMP biopsy images with instant preview and file validation.
Analysis History
Review and restore previous analysis results. Click any history entry to reload its full results.
Export Results
Download complete analysis results — including segmentation overlays — as PNG images.
Architecture
Two-Pipeline System
Classification
DINOv2 + LoRA
1
Input
224 x 224 biopsy image
2
Backbone
DINOv2 ViT-L (700M params, self-supervised)
3
Fine-tuning
LoRA — 18.9M / 700M trainable (2.7%)
4
Ensemble
5-fold cross-validation
5
Augmentation
D4 test-time augmentation (40 predictions)
6
Output
12-class prediction + confidence
11,411 training images
86.25% best fold accuracy
Segmentation
nnU-Net v2
1
Input
Biopsy image (variable size)
2
Architecture
nnU-Net v2, 2D auto-configured
3
Post-processing
Threshold 0.30, hole filling, small object removal
4
Output
Mask, overlay, contour, heatmap views
1,800 + 400 train / val images
IoU: 0.8172
Performance
Honest Metrics, Real Results
0.00%
Best Fold Validation Accuracy
5-fold cross-validation on 11,411 images
0.0000
Segmentation IoU
Validated on 400 images
These metrics represent validation performance. Classification accuracy is the best single fold result from 5-fold cross-validation. Real-world performance may vary depending on image quality and tissue preparation.