Refresh the dataset card

README.md

@@ -61,7 +61,10 @@ configs:
   through the Hub UI, and community pull requests merged into it, are reverted
   by the next publish without warning.
 
-  Change the generator instead:
+  The wording lives in a Markdown template you can edit directly:
+  https://github.com/ImagingDataCommons/idc-index-data/blob/main/scripts/hf/card_template.md
+
+  The counts, tables and front matter around it are generated by:
   https://github.com/ImagingDataCommons/idc-index-data/blob/main/scripts/hf/generate_dataset_card.py
 -->
 
@@ -70,20 +73,17 @@ configs:
 **This dataset is a catalog. It contains metadata and cloud locations for every
 DICOM series in the NCI Imaging Data Commons; it does not contain pixel data.**
 
-[IDC](https://portal.imaging.datacommons.cancer.gov) is an NCI Cancer
-Research Data Commons repository of publicly available cancer imaging data,
-co-located with analysis tools in the cloud. To explore it interactively
-instead, use the [IDC portal](https://portal.imaging.datacommons.cancer.gov/explore/).
-Without downloading anything, any image in IDC can be
-[viewed in the browser](https://learn.canceridc.dev/portal/visualization).
-To query IDC in plain language, point an AI assistant at its
-[agent interfaces](https://learn.canceridc.dev/ai-assistants/agents) --
-a hosted MCP server, an agent skill, and a REST API over the same metadata.
-
-This catalog describes IDC v24 (released 2026-04-16):
-**1,032,911 series** across 166,740 studies,
-85,362 patients and 176 collections,
-totalling **99.3 TB** of imaging data.
+[IDC](https://portal.imaging.datacommons.cancer.gov) is an NCI Cancer Research Data Commons repository of
+publicly available cancer imaging data, co-located with analysis tools in the
+cloud. To explore it interactively instead, use the
+[IDC portal](https://portal.imaging.datacommons.cancer.gov/explore/). Without downloading anything, any image
+in IDC can be [viewed in the browser](https://learn.canceridc.dev/portal/visualization). To query IDC in
+plain language, point an AI assistant at its [agent interfaces](https://learn.canceridc.dev/ai-assistants/agents)
+-- a hosted MCP server, an agent skill, and a REST API over the same metadata.
+
+This catalog describes IDC v24 (released 2026-04-16): **1,032,911 series** across 166,740
+studies, 85,362 patients and 176 collections, totalling
+**99.3 TB** of imaging data.
 
 One row is one DICOM series, with its collection, patient, study and series
 attributes, its license and source DOI, and the S3 URL to fetch it from. Use it
@@ -126,22 +126,22 @@ client.download_from_selection(
 ```
 
 Downloads come directly from IDC's public AWS and GCS buckets at no cost to you.
-What lands on disk is DICOM; read it with [pydicom](https://pydicom.github.io/)
-or [highdicom](https://highdicom.readthedocs.io/).
+What lands on disk is DICOM;
+[Loading images as tensors](#loading-images-as-tensors) below turns it into
+arrays.
 
-Every series in this catalog can also be looked at without downloading
-anything. IDC streams the pixels to a zero-footprint browser viewer, and
-`get_viewer_URL` builds a link to any series you have selected:
+Every series in this catalog can also be looked at without downloading anything.
+IDC streams the pixels to a zero-footprint browser viewer, and `get_viewer_URL`
+builds a link to any series you have selected:
 
 ```python
 print(client.get_viewer_URL(seriesInstanceUID=sel["SeriesInstanceUID"][0]))
 ```
 
-It picks the viewer that fits the data --
-[OHIF](https://github.com/OHIF/Viewers) for radiology,
-[Slim](https://github.com/ImagingDataCommons/slim) for slide microscopy --
-and opens the enclosing study with your series selected. Passing a
-segmentation, as above, brings it up overlaid on the images it segments.
+It picks the viewer that fits the data -- [OHIF](https://github.com/OHIF/Viewers) for radiology,
+[Slim](https://github.com/ImagingDataCommons/slim) for slide microscopy -- and opens the enclosing study with
+your series selected. Passing a segmentation, as above, brings it up overlaid on
+the images it segments.
 
 Query the catalog without downloading anything, using DuckDB:
 
@@ -151,11 +151,189 @@ FROM 'hf://datasets/ImagingDataCommons/idc-index-data/idc_index.parquet'
 GROUP BY 1 ORDER BY size_TB DESC LIMIT 10;
 ```
 
-The same queries run in the **SQL Console** tab on this page, with no local setup.
+The same queries run in the **SQL Console** tab on this page, with no local
+setup.
+
+## Loading images as tensors
+
+A DICOM series is not an array yet. For CT, MR and PET it is usually one file
+per slice. The slices must be ordered by their position in space, not by file
+name, and their stored values rescaled to physical units, such as Hounsfield
+units for CT. Not every series is a volume at all: localizers, uneven slice
+spacing and gantry tilt are all common.
+
+[highdicom](https://highdicom.readthedocs.io/) handles this, and returns a
+`Volume` that keeps voxel spacing and the patient-space affine next to the
+array. The examples below were tested with highdicom 0.28.2.
+
+```bash
+pip install "highdicom>=0.28.2" duckdb idc-index torch
+```
+
+Start in the catalog. `volume_geometry_index` flags every CT, MR and PET series
+whose slices form a regularly spaced 3D grid, so series that won't load as a
+volume are never downloaded:
+
+```python
+import duckdb
+
+hf = "hf://datasets/ImagingDataCommons/idc-index-data"
+query = f"""
+    SELECT SeriesInstanceUID
+    FROM '{hf}/idc_index.parquet'
+    JOIN '{hf}/volume_geometry_index.parquet' USING (SeriesInstanceUID)
+    WHERE collection_id = 'nsclc_radiomics' AND Modality = 'CT'
+      AND regularly_spaced_3d_volume
+    LIMIT 3
+"""
+uids = [row[0] for row in duckdb.sql(query).fetchall()]
+```
+
+Download each series into its own directory, then load it:
+
+```python
+from pathlib import Path
+
+import highdicom as hd
+import numpy as np
+import pydicom
+import torch
+from idc_index import IDCClient
+
+client = IDCClient()
+client.download_from_selection(
+    seriesInstanceUID=uids, downloadDir="idc_data", dirTemplate="%SeriesInstanceUID"
+)
+
+
+def load_volume(uid):
+    files = Path("idc_data", uid).glob("*.dcm")
+    return hd.get_volume_from_series(
+        [pydicom.dcmread(f) for f in files], dtype=np.float32
+    )
+
+
+def volume_to_channel_first_tensor(vol):
+    # Convert to a tensor with a leading channel dimension, as is typically
+    # required in pytorch
+
+    # The result of a match_geometry operation may be permuted/flipped
+    # so the resulting numpy array is non-contiguous
+    arr = np.ascontiguousarray(vol.array)
+
+    if vol.number_of_channel_dimensions == 0:
+        # Volume has no channel -> add one
+        # Result is (channels, slices, rows, columns)
+        t = torch.from_numpy(arr).unsqueeze(0)
+    elif vol.number_of_channel_dimensions == 1:
+        # Volume has a trailing channel -> permute to the front
+        t = torch.from_numpy(arr).permute([3, 0, 1, 2])
+    else:
+        raise ValueError("Expected at most one channel dimension")
+
+    return t
+
+
+vol = load_volume(uids[0])
+image = volume_to_channel_first_tensor(vol)
+
+print(image.shape, vol.spacing)  # spacing in mm, of the three spatial axes
+```
+
+`get_volume_from_series` raises `ValueError` for a series that is not a regular
+grid. Those are the series the geometry filter above leaves out. Pass `dtype`
+explicitly; the default is `float64`.
+
+`get_volume_from_series` has other parameters, for example to choose which pixel
+transforms are applied. See its
+[documentation](https://highdicom.readthedocs.io/en/latest/package.html#highdicom.get_volume_from_series).
+
+### Segmentations
+
+`seg_index` has a row for each segmentation (DICOM SEG) series, and
+`segmented_SeriesInstanceUID` names the image series it segments. This example
+takes one from NLSTSeg, expert segmentations of lung lesions in NLST CT,
+downloads it with its CT, and puts the mask on the CT's grid:
+
+```python
+query = f"""
+    SELECT s.SeriesInstanceUID, s.segmented_SeriesInstanceUID
+    FROM '{hf}/seg_index.parquet' s
+    JOIN '{hf}/idc_index.parquet' i USING (SeriesInstanceUID)
+    JOIN '{hf}/volume_geometry_index.parquet' g
+      ON g.SeriesInstanceUID = s.segmented_SeriesInstanceUID
+    WHERE i.analysis_result_id = 'nlstseg' AND s.total_segments > 1
+      AND g.regularly_spaced_3d_volume
+    LIMIT 1
+"""
+seg_uid, image_uid = duckdb.sql(query).fetchone()
+client.download_from_selection(
+    seriesInstanceUID=[seg_uid, image_uid],
+    downloadDir="idc_data",
+    dirTemplate="%SeriesInstanceUID",
+)
+
+ct = load_volume(image_uid)
+seg = hd.seg.segread(next(Path("idc_data", seg_uid).glob("*.dcm")))
+labels = seg.get_volume(combine_segments=True)
+labels = labels.match_geometry(ct)
+
+image = volume_to_channel_first_tensor(ct)
+mask = volume_to_channel_first_tensor(labels)  # (1, slices, rows, columns)
+print({s.SegmentNumber: s.SegmentLabel for s in seg.SegmentSequence})
+```
+
+A segmentation often covers fewer slices than its image, is rotated relative to
+the source image, or stores the slices in the opposite order; `match_geometry`
+pads, flips and rotates it (as required) onto the image grid. It raises
+`RuntimeError` when the two grids are offset by a fraction of a voxel, and
+rounding in stored positions alone can cause that: some `nsclc_radiomics`
+segmentations sit 2e-5 voxels off their CT, just beyond the default tolerance,
+and `match_geometry(ct, tol=1e-4)` accepts them. Loosen `tol` no further than
+you need: a large tolerance can hide a segmentation that is genuinely misaligned
+with its image.
+
+`combine_segments=True` returns a label map, in which each voxel holds the
+number of the segment it belongs to, or 0. NLSTSeg segments each lesion
+separately, so `mask` numbers the lesions.
+
+A label map cannot hold a voxel that belongs to two segments, and segments may
+overlap. The expert segmentations in `nsclc_radiomics`, for example, outline the
+primary tumor inside the lung that contains it, and `combine_segments=True`
+raises `RuntimeError` on them. For those, `seg.get_volume()` returns one binary
+mask per segment down the last axis, which `volume_to_channel_first_tensor`
+moves to the front: `(segments, slices, rows, columns)`. Losses differ in which
+form they take. PyTorch's `CrossEntropyLoss` takes class indices, as in a label
+map, while others, such as the default behavior of MONAI's `DiceLoss`, take one
+channel per segment. However, be aware that many common loss functions (such as
+Dice) expect the segments to be non-overlapping even when using the
+one-channel-per-segment representation, the so-called "one-hot" format.
+
+`get_volume` can also select a subset of the segments. See its
+[documentation](https://highdicom.readthedocs.io/en/latest/package.html#highdicom.seg.Segmentation.get_volume)
+for this and its other parameters.
+
+### Training and other image types
+
+To feed a `DataLoader`, call `load_volume` from a `torch.utils.data.Dataset`.
+Volumes differ in shape, so bring them to a common one before batching, for
+example with `vol.pad_or_crop_to_spatial_shape((64, 256, 256))` or by
+resampling.
+
+- **Radiographs and mammograms** (CR, DX, MG) are one 2D image per file:
+  `hd.imread(path).get_frame(1)`.
+- **Slide microscopy** (SM) is a multi-resolution pyramid with one file per
+  level, and the full-resolution level is usually too large for one array. Read
+  a region of one level with `hd.imread(path).get_total_pixel_matrix()`, whose
+  `row_start`, `row_end`, `column_start` and `column_end` bounds are 1-based,
+  end excluded. `sm_instance_index` gives each file's `TotalPixelMatrixRows`,
+  `TotalPixelMatrixColumns` and `PixelSpacing_0`, so you can pick the level
+  before downloading.
 
 ## Indices
 
-Each index is a separate config (subset). Load one with the `name` argument of `load_dataset`, or select it from the dropdown in the dataset viewer.
+Each index is a separate config (subset). Load one with the `name` argument of
+`load_dataset`, or select it from the dropdown in the dataset viewer.
 
 | Config | Rows | Size | Description |
 |---|---:|---:|---|
@@ -177,7 +355,7 @@ Each index is a separate config (subset). Load one with the `name` argument of `
 | `volume_geometry_index` | 319,472 | 5.1 MB | This table contains one row per DICOM series from IDC for single-frame CT, MR, and PT SOP classes, with boolean columns characterizing the geometric properties of each series. The checks determine whether the series forms a regularly-spaced rectilinear 3D volume (consistent orientation, spacing, dimensions, and slice positions). Series that do not pass all checks may still be usable with additional processing such as resampling or acquisition geometry correction (e.g., for variable-spacing or gantry-tilted acquisitions). Oblique-aware: uses projection-based slice position computation, which handles gantry-tilted CT, oblique MR, and axial PET uniformly. |
 
 > [!NOTE]
-> `clinical_index` is a *dictionary* of the clinical tables and columns
+> `clinical_index` is a _dictionary_ of the clinical tables and columns
 > available per collection -- not the clinical data itself. The clinical tables
 > are not among these artifacts; retrieve them with
 > `IDCClient.get_clinical_table()`.
@@ -220,7 +398,8 @@ Columns of `idc_index`, the default config:
 | `series_aws_url` | STRING | public AWS S3 URL to download the series in bulk (each instance is a separate file) |
 | `series_size_MB` | FLOAT | total size of the series in megabytes |
 
-Every other config is described by a `<config>_schema.json` sidecar in this repository, carrying the same table and column descriptions:
+Every other config is described by a `<config>_schema.json` sidecar in this
+repository, carrying the same table and column descriptions:
 
 | Config | Columns | Schema |
 |---|---:|---|
@@ -254,7 +433,10 @@ for column in schema["columns"]:
 
 ## Licensing
 
-**The images are not covered by a single license.** Every row carries a `license_short_name` giving the license of that series; the YAML above lists all of them so the dataset appears under each one's Hub filter. Check it per series before redistributing or using data commercially.
+**The images are not covered by a single license.** Every row carries a
+`license_short_name` giving the license of that series; the YAML above lists all
+of them so the dataset appears under each one's Hub filter. Check it per series
+before redistributing or using data commercially.
 
 | License | Series | Commercial use |
 |---|---:|---|
@@ -264,11 +446,16 @@ for column in schema["columns"]:
 | CC BY-NC 3.0 | 5,851 | not allowed |
 | National Library of Medicine Terms and Conditions; May 21, 2019 | 39 | see terms |
 
-Series under *National Library of Medicine Terms and Conditions* are governed by <https://www.nlm.nih.gov/databases/download/terms_and_conditions.html>.
+Series under _National Library of Medicine Terms and Conditions_ are governed by
+<https://www.nlm.nih.gov/databases/download/terms_and_conditions.html>.
 
-Every license IDC uses -- CC BY and CC BY-NC alike -- requires **attribution**. See [IDC licensing and attribution](https://learn.canceridc.dev/data/licensing).
+Every license IDC uses -- CC BY and CC BY-NC alike -- requires **attribution**.
+See [IDC licensing and attribution](https://learn.canceridc.dev/data/licensing).
 
-The index files in this repository are a factual catalog of that content and are distributed under the license of the [idc-index-data repository](https://github.com/ImagingDataCommons/idc-index-data/blob/main/LICENSE). That license covers the tables only, never the referenced images.
+The index files in this repository are a factual catalog of that content and are
+distributed under the license of the
+[idc-index-data repository](https://github.com/ImagingDataCommons/idc-index-data/blob/main/LICENSE). That license
+covers the tables only, never the referenced images.
 
 ## Attribution and citation
 
@@ -277,10 +464,10 @@ Attribution is required by every license in this catalog, and it is owed to the
 the series came from; resolve it to a formatted citation with IDC's citations
 API or `IDCClient.citations_from_selection()`.
 
-Many IDC collections originate from [The Cancer Imaging Archive
-(TCIA)](https://www.cancerimagingarchive.net/); IDC is a TCIA Data Analysis
-Center. Those collections additionally carry TCIA's [data usage policies and
-restrictions](https://www.cancerimagingarchive.net/data-usage-policies-and-restrictions/),
+Many IDC collections originate from
+[The Cancer Imaging Archive (TCIA)](https://www.cancerimagingarchive.net/); IDC
+is a TCIA Data Analysis Center. Those collections additionally carry TCIA's
+[data usage policies and restrictions](https://www.cancerimagingarchive.net/data-usage-policies-and-restrictions/),
 including obligations on downstream attribution.
 
 Please also acknowledge IDC itself:
@@ -314,8 +501,9 @@ version to keep results reproducible:
 load_dataset("ImagingDataCommons/idc-index-data", "idc_index", revision="24.2.2")
 ```
 
-This release, `24.2.2`, indexes IDC v24 (released 2026-04-16). Not every idc-index-data release
-is published here; tags on this repo are a subset of the GitHub releases.
+This release, `24.2.2`, indexes IDC v24 (released 2026-04-16). Not every idc-index-data
+release is published here; tags on this repo are a subset of the GitHub
+releases.
 
 This card is generated, not maintained here. Every publish regenerates
 `README.md` -- YAML front matter and all -- from the release artifacts and
@@ -324,21 +512,28 @@ and community pull requests merged into this card are reverted by the next
 publish, with no warning and no notification to whoever made them. The old text
 survives only in this repo's commit history.
 
-So please don't send card fixes as pull requests here; they will not last.
-Open them against the generator,
-[`scripts/hf/generate_dataset_card.py`](https://github.com/ImagingDataCommons/idc-index-data/blob/main/scripts/hf/generate_dataset_card.py),
+So please don't send card fixes as pull requests here; they will not last. Open
+them against the template the wording comes from,
+[`scripts/hf/card_template.md`](https://github.com/ImagingDataCommons/idc-index-data/blob/main/scripts/hf/card_template.md),
 and they will appear at the next release. Nothing else on the Hub is affected:
 discussions persist, and only the Parquet files, their `*_schema.json` sidecars
 and this card are ever written or removed by the publishing job.
 
 ## Links
 
-- [IDC portal](https://portal.imaging.datacommons.cancer.gov/explore/) -- browse the data and build cohorts interactively
-- [Visualizing IDC images](https://learn.canceridc.dev/portal/visualization) -- how the browser viewers work; get a link to any series with `IDCClient.get_viewer_URL()`
-- [IDC agent interfaces](https://learn.canceridc.dev/ai-assistants/agents) -- search IDC, size a cohort and get a download command by asking: hosted MCP server, agent skill, or REST API
+- [IDC portal](https://portal.imaging.datacommons.cancer.gov/explore/) -- browse the data and build cohorts
+  interactively
+- [Visualizing IDC images](https://learn.canceridc.dev/portal/visualization) -- how the browser viewers
+  work; get a link to any series with `IDCClient.get_viewer_URL()`
+- [IDC agent interfaces](https://learn.canceridc.dev/ai-assistants/agents) -- search IDC, size a cohort and get a
+  download command by asking: hosted MCP server, agent skill, or REST API
 - [IDC documentation](https://learn.canceridc.dev/)
-- [`idc-index` Python package](https://github.com/ImagingDataCommons/idc-index) -- the download client (`pip install idc-index`)
-- [`idc-index-data` on GitHub](https://github.com/ImagingDataCommons/idc-index-data) -- how these tables are built (SQL included)
+- [`idc-index` Python package](https://github.com/ImagingDataCommons/idc-index)
+  -- the download client (`pip install idc-index`)
+- [`idc-index-data` on GitHub](https://github.com/ImagingDataCommons/idc-index-data) -- how these tables are built
+  (SQL included)
+- [`highdicom` Python package](https://github.com/ImagingDataCommons/highdicom)
+  -- used to read DICOM and arrange as tensors
 - [GCS mirror of the release artifacts](https://storage.googleapis.com/idc-index-data-artifacts?prefix=current/release_artifacts/)
   -- fetch a single file directly, e.g.
   `https://storage.googleapis.com/idc-index-data-artifacts/current/release_artifacts/idc_index.parquet`