/autoPETIII

Final submission for the autoPET III challenge.

Primary LanguagePythonMIT LicenseMIT

AutoPET III Challenge: Final Submission

This repository contains the code and models used for the final submission in the AutoPET III Challenge. This repository is released under the MIT License.

Overview

Our method uses a classifier to differentiate between FDG and PSMA tracers. It then runs inference on the PET/CT using a tracer-specific nnU-Net ensemble. The paper is available at: https://arxiv.org/pdf/2409.12155

Segmentation Models:

  • FDG Model: nnUNet ensemble specifically trained with FDG PET data.
  • PSMA Models: Includes two models trained on PSMA PET data:
    1. A standard nnU-Net architecture.
    2. A nnU-Net model with a Residual Encoder architecture.

Classifier:

  • Tracer Classifier: A model trained to classify the input as either FDG or PSMA tracer. This classifier can be used if the used tracer is unknown.

Model Checkpoints

All model weights are available under https://drive.google.com/file/d/1nY7ciiJPcfxtv1XFpY-eWsmBkxfTSJez/view. They include the following files/folders:

  • FDG Model: Checkpoints are found in the Dataset001_fdgweighted folder.
  • PSMA Models:
    • Standard nnU-Net: Located in Dataset002_psmaweighted.
    • Residual Encoder nnU-Net: Located in Dataset003_psmaweighted.
  • Tracer Classifier: Model weights for the tracer classifier are available in tracer_classifier.pt.