Challenge data

Description

Competitive challenge

Biology

Health

Segmentation

Images

10MB to 1GB

Advanced level

Dates

Started on Jan. 8, 2025

Challenge context

CT scans provide highly accurate 3D images of the human body (up to 0.5 mm resolution), enabling us to grasp the human anatomy.
The aim of this challenge is to automatically segment the anatomical structures of the human body, as well as tumors, on a CT scan, without semantically identifying the associated organs. In other words, the aim is to identify the visible shapes on a CT scan.
This object discovery problem is an intuitive and straightforward one for a human, who will be able to identify new objects on a scene very easily, and even on a CT-scan, even if he or she has never seen the object before.
In the image below, from an abdominal CT-scan, the various structures have been segmented:

![](/media/public/raidium_2024_1.png)

In addition, numerous public datasets are available, with annotations for specific organs. 
It may seem desirable to merge many datasets to learn a more robust segmentation model.
However, semantic segmentation algorithms don't work well when datasets don't have the same labels: 
for example, the liver will be indicated as “background” in one dataset, and labeled in another, which will disrupt model learning.

Challenge goals

The goal is to **segment structures using their shape, but no exhaustive annotations**.
The training data is composed of two types of images:

* Partially annotated CT images with anatomical segmentation masks of individual structures
    * They act as the ground truth definition of what are anatomical structures. 
    * However, they don’t intend to be representative of all of the possible structures and their diversity, but can still be used as training material.
    * The masks do not contain all the annotated organs in the entire dataset. For example, on two abdominal images, 
	the mask for A will contain the liver and spleen, while the mask for B will contain only the spleen (even though the liver is visible on the image).

* Raw CT images, without any segmented structures
    * They can be used as additional training material, in an unsupervised setting.

The test set is made of new images with all their corresponding segmented structures, and the metric measures the capacity to correctly segment and separate the different structures on an image.

Data description

The input is a list of 2D grayscale images (i.e. a 3D numpy array), each corresponding to a slice of a CT-scan (in the transverse plane) of size 256x256 pixels. The slices are randomized and thus there is no 3D information.

The label/output is a list of 2D matrices (i.e a 3D numpy array) of size 256x256 pixels, with integer (uint8) values. Each position (w,h) of each matrix $Y\_{i,w,h}$ identifies a structure.

For example on the Figure 1 above, the  colors correspond to different segmented structures, thus each pixel label $Y_{i,w,h}$ at position (w,h) has values in the integer range `[[0; 23]]`. 0 is a special value meaning this pixel is not part of any structure and then part of the background.

In practice, the output is encoded as a CSV file that encodes the **transpose** of the **flattened** label matrix.

Note: the transpose is used here for performance reasons. Pandas is very slow to load CSV files with many columns, but is very fast to load CSV files with many rows.

Thus, the CSV is composed of 65 536 (256x256) columns, each corresponding to a pixel of the image, and N rows, each corresponding to an image.

To get the list of 2D predictions, you must therefore transpose the received CSV, and reshape it:
```python
import pandas as pd
predictions = pd.read_csv(label_csv_path, index_col=0, header=0).T.values.reshape((-1, 256, 256))
# In the end, we get a list of 2D predictions, i.e. 3D numpy array of shape (N, 256, 256)
```

In order to get the output CSV from a list of 2D predictions, it is necessary to flatten each of the predictions, and concatenate them into a single matrix, then transpose the matrix and finally save it in a CSV file:
```python
import pandas as pd
# predictions is a list of 2D predictions, i.e. a 3D numpy array of shape (500, 256, 256)
predictions = np.array of [prediction_1, prediction_2, prediction_3, ...]
pd.DataFrame(predictions.reshape((predictions.shape[0], -1))).T.to_csv(output_csv_path)
```

This problem can be seen as an image-wise pixel clustering problem, where each structure is a cluster in the image: the pixel labels values are structure identifiers on a specific image and are not necessarily coherent between images. For example the structure associated with the liver can be mapped to the label 4 on one image and to 1 on another image.

The train set is composed 2000 images, split into two groups:

* 800 with fully segmented structures. 
    * For these images, the corresponding label is a 2D matrix with pixel labels for segmented structures, and the other pixels set to 0
* 1200 of them have no annotations at all.
    * For these images, the corresponding label is a 2D matrix filled with zeros

The data also includes a file: `annotated_labels.json`, which indicates which organs are annotated for each scan.

These organs can be present or absent in the segmentation mask. If they are absent, this indicates with certainty that the organ is *not* present in this image.

On the contrary, for a given image, if a label is not present in `annotated_labels`, this indicates that it can be present or absent in the image, without additional information.

You can use this file:
- to calculate your metrics on a validation set (and not penalize false positives in case of unannotated organ)
- for your training, in order to use only the annotated labels

Note: segmentation maps of structures of the train set, in the form of images (in addition to the CSV) are given in the supplementary materials.

The test set is composed of 500 images with organ and tumor segmentation. For these images, the corresponding label is a 2D matrix with the segmented structures, and the other pixels set to 0.
Considering the fact that an individual image with its label matrix is around 400KB, and we have 500 images, we then have a dataset of about 200MB in total.

Note: the segmentation map is not dense, meaning that some pixels between structures are not segmented, as we can see on the image above. These pixels are considered to be part of the background.

Note: it is not authorized to use additional radiological training data, or radiological pre-trained models, or any other external radiological data source. You are however allowed to use pre-trained models and data that are not radiological (DINO v2, SAM ...).

The global metric is computed by averaging the DICE score for each organ, then each image.

# Notebook

A getting started notebook can be found at the following address:

[https://colab.research.google.com/drive/1tjL_vct6dihVApl8oBVg6_6JFk-U7F5S#scrollTo=EZSMQF6uSJ4t](https://colab.research.google.com/drive/1tjL_vct6dihVApl8oBVg6_6JFk-U7F5S#scrollTo=EZSMQF6uSJ4t)

It includes examples of data loading, visualization, metric computation and a baseline.

Benchmark description

Files

Files are accessible when logged in and registered to the challenge

The challenge provider

Pierre Manceron - Head of Science at Raidium

PROVIDER WEBSITE

Challenge Data

Anatomic structures segmentation with missing labels
by Raidium

Description

Dates

Challenge context

Challenge goals

Data description

Benchmark description

Files

The challenge provider

Challenge Data

Anatomic structures segmentation with missing labels by Raidium

Description

Dates

Challenge context

Challenge goals

Data description

Benchmark description

Files

The challenge provider

Anatomic structures segmentation with missing labels
by Raidium