Automated Head and Neck Tumor Segmentation from 3D PET/CT HECKTOR 2022 Challenge Report

Andriy Myronenko, Md Mahfuzur Rahman Siddiquee, Dong Yang, Yufan He, Daguang Xu

TL;DR

This paper presents the winning solution for the HECKTOR 2022 challenge, which focuses on automating the segmentation of head and neck tumors and lymph nodes from 3D PET/CT scans. The team achieved first place by employing a robust pipeline involving image resampling, cropping, and a SegResNet model trained with deep supervision, ultimately outperforming other participants with a high Dice score.

What Problem Does This Paper Address?

Head and neck (H&N) cancer is a significant global health concern, and accurate segmentation of tumors and lymph nodes from medical imaging (CT and PET) is crucial for effective treatment planning, particularly for radiotherapy. However, manual segmentation is time-consuming and prone to inter-observer variability. The HECKTOR 2022 challenge aimed to drive the development of automated 3D segmentation algorithms for these H&N tumors (GTVp) and lymph nodes (GTVn) using 3D PET/CT data. The challenge presented a complex dataset with varying resolutions and large imaging volumes, making automated segmentation a demanding task.

Methods

The authors' winning approach, developed by the NVAUTO team, is built upon the MONAI framework and leverages the Auto3DSeg system for automated parameter selection. The core of their segmentation model is SegResNet, an encoder-decoder architecture incorporating deep supervision.

Key methodological steps include:

  • Image Preprocessing: All CT and PET images were resampled to a common isotropic resolution of 1x1x1mm. A crucial step was cropping the images to focus on the head and neck region, reducing computational load and irrelevant anatomical information. This cropping was achieved using a relatively simple heuristic based on PET thresholding and centerline detection, which proved highly effective.
  • Data Normalization: CT images were normalized to a 0-1 interval, while PET images were normalized to zero mean and unit standard deviation. Both normalized images were then concatenated to form a 2-channel input.
  • Model Architecture: SegResNet, a ResNet-based encoder-decoder network with deep supervision, was employed. Deep supervision involves adding auxiliary loss functions at intermediate layers of the decoder, which helps to improve gradient flow and performance, especially in deeper networks.
  • Training Strategy: The model was trained using a 5-fold cross-validation approach. Data augmentation techniques, including random affine transformations, flips, and intensity augmentations (scale, shift, noise, blurring) for CT, were applied to improve model robustness.
  • Loss Function: A combined Dice and CrossEntropy loss was used, summed across all deep supervision sublevels.
  • Ensembling: The final submission was an ensemble of 15 models derived from three fully trained runs. Test-time augmentation (TTA) with axis flips was also employed, which significantly boosted performance.

Datasets

The primary dataset used is the HECKTOR 2022 dataset. This dataset comprises 883 patients with 3D CT and PET scans, where 524 cases were provided with ground truth labels for GTVp and GTVn. The data originates from multiple institutions (9 centers across Canada, Switzerland, France, and the United States). The CT scans had an average resolution of 0.98x0.98x3 mm, and PET scans had an average resolution of 4x4x4 mm. The challenge data is notable for its large imaging volume (up to 500x500x1000 mm) and the fact that the tumor region constitutes a small fraction of the total image volume.

Key Insights and Ideas

  • Effective Cropping is Key: A simple, anatomically-guided cropping strategy significantly improved efficiency and performance by focusing the model on the relevant H&N region, avoiding the need for more complex deep-learning-based cropping.
  • Deep Supervision Enhances Performance: The use of deep supervision in the SegResNet architecture proved beneficial for training a robust segmentation model, likely by providing better gradient signals throughout the network.
  • Ensembling and TTA are Powerful: Combining multiple trained models through ensembling and applying test-time augmentation (TTA) were critical for achieving top performance, demonstrating their continued importance in medical image analysis challenges.
  • MONAI and Auto3DSeg Streamline Development: The use of the MONAI framework and Auto3DSeg system facilitated rapid development and experimentation, allowing the team to focus on the core segmentation task.

Key Results

The NVAUTO team achieved 1st place in the HECKTOR 2022 challenge. Their best submission, which utilized an ensemble of models with Test Time Augmentation (TTA), achieved an aggregated Dice score of 0.78802. This performance placed them at the top of the HECKTOR22 challenge leaderboard, with their three submitted entries occupying the top three positions.

Limitations and Future Work

  • Limitations: The paper mentions that their simple cropping heuristic had 100% success on the training set, but it's possible that more complex cases or unseen data might pose challenges. The "post-processing" submission, which aimed to refine lymph node segmentation by removing small components, was ultimately not helpful and reduced accuracy, suggesting that simpler, end-to-end learned approaches are often superior to heuristic post-processing.
  • Future Work: While not explicitly detailed, the success of their automated pipeline suggests that further exploration of advanced augmentation strategies, different ensemble techniques, or potentially more sophisticated network architectures could lead to even higher performance in future challenges. The authors also note that the challenge included an outcome prediction task, which was not addressed in this report, leaving room for future work in that area.

Summary updated

Abstract

Head and neck tumor segmentation challenge (HECKTOR) 2022 offers a platform for researchers to compare their solutions to segmentation of tumors and lymph nodes from 3D CT and PET images. In this work, we describe our solution to HECKTOR 2022 segmentation task. We re-sample all images to a common resolution, crop around head and neck region, and train SegResNet semantic segmentation network from MONAI. We use 5-fold cross validation to select best model checkpoint. The final submission is an ensemble of 15 models from 3 runs. Our solution (team name NVAUTO) achieves the 1st place on the HECKTOR22 challenge leaderboard with an aggregated dice score of 0.78802 (https://hecktor.grand-challenge.org/evaluation/segmentation/leaderboard/.). It is implemented with Auto3DSeg (https://monai.io/apps/auto3dseg.).
BibTeX
@inproceedings{myronenko2022automated,
  author = {Andriy Myronenko and Md Mahfuzur Rahman Siddiquee and Dong Yang and Yufan He and Daguang Xu},
  title = {{Automated Head and Neck Tumor Segmentation from 3D PET/CT HECKTOR 2022 Challenge Report}},
  booktitle = {HECKTOR @ MICCAI},
  year = {2022},
  doi = {10.1007/978-3-031-27420-6\_2},
}
Figure