You are a computational pathologist. You have been given a digitized whole-slide image (WSI) of a sentinel lymph node biopsy from a breast cancer patient. Your task is to examine the tissue and identify all regions containing metastatic tumor cells.

== CLINICAL CONTEXT ==

In breast cancer surgery, sentinel lymph nodes (the first lymph nodes draining the tumor) are biopsied and examined by pathologists for metastases. The presence and size of metastases determines the pathological N-stage, which directly impacts treatment decisions.

Tumor deposits in lymph nodes appear as clusters of epithelial cancer cells that are morphologically distinct from the normal lymphoid tissue. They often show:
- Irregular glandular or sheet-like architecture
- Large, pleomorphic nuclei with prominent nucleoli
- Increased nuclear-to-cytoplasmic ratio
- Pale or eosinophilic cytoplasm
- Sometimes desmoplastic stromal reaction

Normal lymph node tissue contains lymphocytes (small, dark, round nuclei), germinal centers (pale zones with mixed cells), sinuses (open spaces with histiocytes), and fibrous capsule/trabeculae.

== AVAILABLE TOOLS ==

The WSI is in TIFF format and is a multi-resolution image pyramid (gigapixel at full resolution). You CANNOT load the entire image into memory. Instead, use the OpenSlide Python library:

    pip install openslide-python

Key API:
    import openslide
    slide = openslide.OpenSlide("input/slide.tif")
    slide.dimensions          # (width, height) at level 0
    slide.level_count         # number of pyramid levels
    slide.level_dimensions    # dimensions at each level
    slide.level_downsamples   # downsample factor per level
    slide.read_region((x, y), level, (w, h))  # read a patch
    slide.get_thumbnail((w, h))               # low-res overview

IMPORTANT: The location parameter in read_region is ALWAYS in level-0 (full resolution) coordinates, regardless of which level you read from.

== INPUT FILES ==

input/slide.tif              - The whole-slide image (TIFF, multi-resolution pyramid)
input/slide_metadata.json    - Metadata including:
    slide_id                 - Identifier for this slide
    tissue_type              - "sentinel_lymph_node"
    stain                    - "H&E"
    microns_per_pixel        - Physical pixel size at level 0
    analysis_tile_size       - Tile size in pixels for the analysis grid (256)
    analysis_level           - Which pyramid level to use for grid definition (0)

== YOUR TASK ==

1. Open the slide and understand its structure (dimensions, levels, magnification)
2. Generate a low-resolution overview to see where tissue is located
3. Identify tissue-containing regions (distinguish tissue from empty glass background)
4. Systematically scan tissue regions, starting at low magnification and zooming into suspicious areas
5. Classify each analysis grid tile as tumor-positive or tumor-negative
6. The analysis grid divides the slide into non-overlapping tiles of size analysis_tile_size × analysis_tile_size pixels at level analysis_level (level 0). Grid tile (x, y) corresponds to the region starting at pixel (x * tile_size, y * tile_size).

== REQUIRED OUTPUT ==

Write a single JSON file to output/predictions.json with the following structure:

{
    "slide_id": "<from slide_metadata.json>",
    "tumor_present": true or false,
    "tumor_tiles": [
        {"x": <int>, "y": <int>},
        ...
    ],
    "tumor_tile_count": <int>,
    "total_tissue_tiles": <int>,
    "tumor_area_fraction": <float>
}

Where:
- tumor_present: true if you identified any tumor tiles, false otherwise
- tumor_tiles: list of (x, y) grid coordinates of tiles containing tumor
- tumor_tile_count: length of tumor_tiles list
- total_tissue_tiles: number of grid tiles containing tissue (not background)
- tumor_area_fraction: tumor_tile_count / total_tissue_tiles

== STRATEGY HINTS ==

- Start with a low-resolution overview (level 2 or 3) to identify tissue regions
- Use intermediate resolution (level 1) to scan for suspicious areas
- Zoom to full resolution (level 0) only to confirm tumor in specific tiles
- Tissue vs. background: tissue is pink/purple, background is white. A simple approach is Otsu thresholding on grayscale.
- Do NOT try to read the entire slide at level 0 — it will crash. Work tile-by-tile.