# CPAISD V12 Final Report

Date: 2026-06-04

## Final Model Position

The final submitted model is V12, a high-resolution 2.5D supervised U-Net-style segmentation model for NCCT ischemic-area segmentation. It is not reported as a DINOv3/3DINO foundation-model result. The primary output is ischemic area, defined as core plus penumbra. A separately trained V12 core-only model is included as an exploratory auxiliary model.

## Main Test Results

| Model | Target | Threshold | Dice | IoU | Precision | Recall |
|---|---|---:|---:|---:|---:|---:|
| V12_penumbra | core + penumbra ischemic area | 0.55 | 0.4217 | 0.2672 | 0.5505 | 0.3418 |
| V12_core | core-only mask | 0.70 | 0.0701 | 0.0363 | 0.0734 | 0.0670 |

## Key Parameters

| Parameter | V12_penumbra | V12_core |
|---|---:|---:|
| Image size | 384 x 384 | 384 x 384 |
| Context slices | 5 | 5 |
| Folds | 5 | 5 |
| Epochs | 35 | 35 |
| Batch size | 2 | 2 |
| Learning rate | 0.0001 | 0.0001 |
| Weight decay | 0.0005 | 0.0005 |
| U-Net base channels | 24 | 24 |
| Dropout2d | 0.08 | 0.08 |
| Early stopping patience | 7 | 7 |
| Tissue HU range | 18.0 to 90.0 | 18.0 to 90.0 |
| Tversky alpha FP / beta FN | 0.75 / 0.25 | 0.75 / 0.25 |
| BCE positive weight | 2.0 | 2.0 |

## Fold Validation

V12_penumbra mean best validation Dice: 0.3715.  
V12_core mean best validation Dice: 0.2753.

## Interpretation

V12_penumbra is the final primary model because it achieved the strongest held-out test performance and the predictions are clinically more interpretable for the full ischemic region. V12_core is much harder because core labels are smaller, more imbalanced, and less visually separable on NCCT; it should be presented as exploratory rather than the primary endpoint.

## Viewer

Cloudflare Pages viewer: https://32c21435.cpaisd-v11-v12-viewers.pages.dev