PDAC tumor segmentation on T1W-CE-MRI :PANTHER challenge baseline


Logo for PDAC tumor segmentation on T1W-CE-MRI :PANTHER challenge baseline

About

Creator:

User Mugshot amparobt9 

Image Version:
d10e3a11-1da0-49c5-867c-00b826ff57c1
Last updated:
April 13, 2025, 10:52 p.m.
Model Version:
1bb07697-6c8f-4570-8ce0-0804b700c9ee
Last updated:
April 13, 2025, 7:07 p.m.

Interfaces

This algorithm implements all of the following input-output combinations:

Inputs Outputs
1
    Abdominal T1 MRI
    Pancreatic tumor segmentation

Challenge Performance

Date Challenge Phase Rank
April 16, 2025 PANTHER Open Development Phase Task 1 1

Model Facts

Summary

This is the baseline algorithm for Task 1 of the PANTHER challenge. The algorithm is built on the nnU-Net framework (v2) and employs a two-step process to segment pancreatic ductal adenocarcinoma (PDAC) on T1-weighted, contrast-enhanced arterial phase MRIs. In the first step, a nnU-Net model is trained for tumor segmentation. In the second step, the segmentation is refined through post-processing: the input image is downsampled to half its resolution, and a pancreas mask is generated using MRSegmentator. The generated pancreas mask is then upsampled to the original resolution, and connected component analysis is applied to the tumor segmentation.

Finally, only components that contain at least 10 voxels and have an overlap of at least 15% with the pancreas mask are retained, while the others are discarded. This process is summarized in the figure below.

Figure: Pipeline of the baseline algorithm for Task 1 of the PANTHER challenge.

Mechanism

This algorithm automatically generates a segmentation mask for pancreatic ductal adenocarcinoma (PDAC) on T1-weighted, contrast-enhanced arterial phase MR images.

  • Input: T1-weighted, contrast-enhanced arterial phase MRI
  • Output: Segmentation mask delineating PDAC regions

The model was trained using the 92 annotated MR images from the PANTHER Public Training Dataset.

Validation and Performance

Uses and Directions

Purpose:
This algorithm was developed for research purposes only and is designed to automatically segment pancreatic ductal adenocarcinoma (PDAC) on diagnostic T1-weighted contrast-enhanced arterial phase MR images.

Intended Use:
- The model is intended exclusively for use on arterial phase T1W CE MR images.
- It should only be applied to image data that match the imaging characteristics of the training set.

Benefits:
- Automatic segmentation of PDAC lesions

Target Population:
This algorithm was trained on a cohort of patients with confirmed or suspected PDAC. It is meant to support radiologists evaluating pancreatic tumors using arterial phase MR images and is not intended for use on other imaging modalities or sequences.

General Use:
Radiologists can use this model as a decision support tool in the analysis of diagnostic MR images. The segmentation output should be interpreted together with other clinical data. In practice, if the model highlights a region as PDAC, the radiologist should review the finding in the context of the patient’s overall clinical picture.

Before Using This Model:
It is recommended to validate the model retrospectively and prospectively on a cohort that reflects your local patient population to confirm its performance and clinical applicability.

Safety and Efficacy Evaluation:
To be determined.

Warnings

Risks:
Even when used appropriately, reliance on this PDAC segmentation algorithm can result in misdiagnosis due to incorrect predictions or missed regions of tumor. It's important to note that no algorithm is perfect, and errors in segmentation can adversely affect decision-making.

Inappropriate Settings:
This model was developed exclusively for diagnostic T1-weighted contrast-enhanced arterial phase MR images. It was not trained on cases with significant artifacts, hypervascular or isovascular tumors. Consequently, if the algorithm is applied to such images, it may produce unreliable predictions. It should not be used in clinical practice without thorough local evaluation.

Generalizability:
Since the model was trained on a specific set of diagnostic MR images, its performance may diminish when applied to images from different sources or acquired using different protocols. It is critical to validate the model on the intended patient population before it is applied in an external setting.

Common Error Messages

The inference progress for this algorithm is incorrectly captured as "error messages", resulting in each case being predicted with "Succeeded, with warnings". This does not mean actual warnings occurred, although this can still be the case.

Information on this algorithm has been provided by the Algorithm Editors, following the Model Facts labels guidelines from Sendak, M.P., Gao, M., Brajer, N. et al. Presenting machine learning model information to clinical end users with model facts labels. npj Digit. Med. 3, 41 (2020). 10.1038/s41746-020-0253-3