# UA-HFNet — Project Structure

## Step 0: Run data inspection first
```bash
python scripts/inspect_data.py --data_root /your/data/path
```

Before running, open `scripts/inspect_data.py` and edit the `CENTER_CONFIG`
dict at the top to match your actual filenames. The key things to set:

| Variable | What to fill |
|---|---|
| `"center1_CT.npy"` | your actual CT filename for center 1 |
| `"center1_labels.csv"` | your CSV filename for center 1 |
| `"center23_labels.csv"` | your merged CSV for centers 2+3 |
| `csv_filter` | set to `None` if the CSV has no center_id column |

## What the script checks
- Shape of every .npy file and whether N matches the CSV row count
- Value range (min/max/mean/std) per modality — helps verify normalization
- NaN and Inf counts
- Label distribution (csPCa prevalence per center)
- Whether total N == 527

## After running, tell me:
1. The shape output for each modality (e.g. `(363, 32, 64, 64)`)
2. The value ranges — especially PET (should be SUV ~0–20 before norm)
3. Any warnings or errors

Then I will write:
- `data/dataset.py`   — PyTorch Dataset class
- `data/preprocess.py` — resize / normalize if needed
- `data/augment.py`   — 3D augmentations
- `configs/config.yaml` — all hyperparameters

## Project layout (will be built incrementally)
```
ua_hfnet/
├── scripts/
│   └── inspect_data.py     ← START HERE
├── data/
│   ├── dataset.py
│   ├── preprocess.py
│   └── augment.py
├── models/
│   ├── encoder.py          (3D ResNet-18)
│   ├── pillar1_edl.py      (uncertainty heads)
│   ├── pillar2_attn.py     (cross-modal attention)
│   └── ua_hfnet.py         (full model)
├── losses/
│   ├── focal_loss.py
│   └── edl_loss.py
├── train/
│   ├── train.py
│   ├── evaluate.py
│   └── cross_val.py
├── baselines/
│   ├── concat_model.py
│   └── sota_models.py
├── viz/
│   ├── attn_map.py
│   ├── uncertainty_viz.py
│   └── results_plot.py
├── utils/
│   └── metrics.py
└── configs/
    └── config.yaml
```