T10: Segmenting lesions within ROIs in CT¶

Objective:
Develop a model to segment 3D lesion masks within volumes-of-interest (VOIs) from CT scans. Each VOI is centered on a known lesion voxel, simulating a radiologist’s click.

Patient Population:

725 cases sourced from Radboudumc and Jeroen Bosch Ziekenhuis in the Netherlands.
- 46.5% female
- Each case contains a single lesion within the provided VOI

Imaging Data:

• 3D VOIs of size 256×256×128 voxels centered on the lesion
• Where needed, volumes were padded using a value of (minimum intensity – 1)
• Each VOI contains a single annotated lesion.
• Each lesion was segmented independently by three annotators; majority vote masks were reviewed and corrected by an expert radiologist

Test Data:
Unlabeled VOIs with the same dimensions and characteristics as the training data. Participants must predict a binary segmentation mask for the lesion in each VOI.

Reference Standard:

• Consensus lesion masks derived from three annotators via majority vote
• Final masks reviewed and refined by an experienced radiologist

Evaluation Metrics:

• Segmentation Performance (SP): Sørensen-Dice coefficient
• Long-Axis Error (LAE): Symmetric mean absolute percentage error in the lesion’s long axis
• Short-Axis Error (SAE): Symmetric mean absolute percentage error in the lesion’s short axis

A composite score (CS) is calculated as:
CS = 0.888 · SP + 0.056 · LAE + 0.056 · SAE

Relation to Existing Challenges:¶

• Task 10 uses the same test dataset as the ULS Challenge,
• However, it simplifies evaluation by excluding the segmentation consistency score from the composite metric, focusing instead on individual accuracy. Moreover, the original ULS evaluation computes the exact longest and perpendicular short diameters by extracting boundary points and brute-forcing all pairwise distances (and then finding the true perpendicular), which is more accurate but computationally intensive. This evaluation step was simplified by fitting an ellipse to each 2D slice using regionprops and takes its major/minor axes (and centroids ± half-axes), making it faster and simpler but only provides an approximation of the true lesion shape.