RSNA 2024 Lumbar Spine Degenerative Classification¶

Modality: MRI | Format: DICOM (per-series volumes) | Dim: 3D | Labels: 75 severity columns

Overview¶

The RSNA 2024 Lumbar Spine Degenerative Classification challenge dataset contains MRI studies of the lumbar spine labelled for five degenerative conditions at five lumbar levels (L1/L2 through L5/S1), giving 25 condition–level combinations. Each is labelled Normal/Mild, Moderate, or Severe — encoded as 75 one-hot severity columns (3 per condition-level pair). Three MRI series types are available per study: Axial T2, Sagittal T1, and Sagittal T2/STIR.

Download¶

Available on Kaggle: RSNA 2024 Lumbar Spine. Requires Kaggle account.

Expected layout¶

rsna-2024-lumbar-spine-degenerative-classification/
  train.csv
  train_series_descriptions.csv
  train_label_coordinates.csv     # point annotations (optional)
  train_images/
    <study_id>/
      <series_id>/
        <instance_number>.dcm

Label columns¶

75 columns in the format <condition>_<level>_<severity>:

Conditions: spinal_canal_stenosis, left_neural_foraminal_narrowing, right_neural_foraminal_narrowing, left_subarticular_stenosis, right_subarticular_stenosis

Levels: l1_l2, l2_l3, l3_l4, l4_l5, l5_s1

Severities: normal_mild, moderate, severe

Example: spinal_canal_stenosis_l1_l2_normal_mild, spinal_canal_stenosis_l1_l2_moderate, spinal_canal_stenosis_l1_l2_severe

Constructor arguments¶

Argument	Type	Required	Default	Description
`base_image_dir`	`str`	Yes	`None`	`train_images/` (or `test_images/` for the test split) — direct parent of `<study_id>/<series_id>/` (e.g. `~/datasets/competitions/rsna-2024-lumbar-spine-degenerative-classification/train_images/`)
`csv_path`	`str`	No	auto	`train.csv`; auto-discovered
`series_description_csv_path`	`str`	No	auto	`train_series_descriptions.csv`
`coord_csv_path`	`str`	No	auto	`train_label_coordinates.csv` for point bbox annotations
`series_filter`	`str\\|None`	No	`None`	Load only one series type: `"Axial T2"`, `"Sagittal T1"`, or `"Sagittal T2/STIR"`

Shared arguments (inherited from `BaseRadiologicalDataset`)¶

Argument	Type	Required	Default	Description
`output_cls`	`bool`	No	`False`	Include `"cls"` tensor in data dict
`output_mask`	`bool`	No	`False`	Include `"mask"` in data dict
`output_report`	`bool`	No	`False`	Include `"report"` in data dict
`output_bbox`	`bool`	No	`False`	Include `"bbox"` and `"bbox_labels"` in data dict
`transform`	MONAI transform	No	`None`	MONAI Compose transform; `None` uses the default pipeline
`cache_dir`	`str`	No	`"./cache"`	MONAI cache directory
`dtype`	`torch.dtype`	No	`torch.bfloat16`	Output tensor dtype
`harmonized_df`	`pd.DataFrame`	No	`None`	Pre-built harmonized DataFrame
`harmonizer`	harmonizer	No	`None`	Pre-instantiated harmonizer
`harmonizer_path`	`str`	No	`None`	Path to saved harmonized CSV

Dataset constructor¶

The constructor arguments above apply to both RSNA2024LumbarSpineTrainDataset and RSNA2024LumbarSpineTestDataset. The test split has no labels — output_cls and output_bbox are silently ignored, and only series_description_csv_path is needed (no train.csv or coords CSV).

Train split¶

import torch
from radharmony.dataset import RSNA2024LumbarSpineTrainDataset

# Load Sagittal T2/STIR series only
ds = RSNA2024LumbarSpineTrainDataset(
    base_image_dir="/data/rsna-2024-lumbar-spine/train_images/",
    series_filter="Sagittal T2/STIR",
    output_cls=True,
    dtype=torch.float32,
)
train_ds, val_ds = ds.get_datasets(n_splits=5)

Test split¶

from radharmony.dataset import RSNA2024LumbarSpineTestDataset

ds = RSNA2024LumbarSpineTestDataset(
    base_image_dir="/data/rsna-2024-lumbar-spine/test_images/",
    series_filter="Sagittal T2/STIR",
    dtype=torch.float32,
)

Harmonizer¶

from radharmony.harmonizer import RSNA2024LumbarSpineTrainHarmonizer

h = RSNA2024LumbarSpineTrainHarmonizer(
    csv_path="/data/rsna-2024-lumbar-spine/train.csv",
    base_image_dir="/data/rsna-2024-lumbar-spine/train_images/",
    series_description_csv_path="/data/rsna-2024-lumbar-spine/train_series_descriptions.csv",
)
df = h.harmonize()
print(df.columns.tolist())
df.to_csv("rsna_lumbar_harmonized.csv", index=False)

Load from saved harmonized CSV¶

import pandas as pd
from radharmony.dataset import RSNA2024LumbarSpineTrainDataset

ds = RSNA2024LumbarSpineTrainDataset(
    base_image_dir="/data/rsna-2024-lumbar-spine/train_images/",
    harmonized_df=pd.read_csv("rsna_lumbar_harmonized.csv"),
    output_cls=True,
)

Harmonizer notes¶

image_path points to a series directory; ITKReader loads it as a 3D volume
Labels are one-hot encoded across three severity levels per condition-level pair
series_filter is applied at load time to select a subset of series types
Point annotations from train_label_coordinates.csv are represented as small dot bounding boxes when output_bbox=True
Each study may contribute up to 3 rows (one per series type) unless filtered

Outputs¶

Flag	Key	Shape	Notes
`output_cls=True`	`"cls"`	`(75,)`	One-hot severity across 25 condition-level pairs
`output_bbox=True`	`"bbox"`, `"bbox_labels"`	list	Point annotations as dot boxes

Example paths¶

Role	Path
`base_image_dir` (train)	`/path/to/rsna-2024-lumbar-spine-degenerative-classification/train_images/`
`base_image_dir` (test)	`/path/to/rsna-2024-lumbar-spine-degenerative-classification/test_images/`
`csv_path` (`train.csv`)	`/path/to/rsna-2024-lumbar-spine-degenerative-classification/train.csv`
`series_desc_csv_path` (`train_series_descriptions.csv`)	`/path/to/rsna-2024-lumbar-spine-degenerative-classification/train_series_descriptions.csv`
`label_coord_csv_path` (`train_label_coordinates.csv`)	`/path/to/rsna-2024-lumbar-spine-degenerative-classification/train_label_coordinates.csv`