LNQ Challenge 2023: Learning Mediastinal Lymph Node Segmentation with a Probabilistic Lymph Node Atlas

TL;DR

Mediastinal lymph node segmentation on CT is hindered by poor contrast, variable node morphology, and partial annotations. The paper introduces a probabilistic lymph node atlas and atlas-to-patient distance priors, integrated into an enhanced nnU-Net with PA-weighted loss and post-processing, plus strong, domain-generalizing augmentations and oversampling of fully annotated data. Ablation shows priors alone provide limited gains, while combining priors with full annotations and augmentation yields the largest improvements; semi-supervised attempts did not improve performance. The best 5-model ensemble with post-processing achieves a Dice score of $0.6033$ and identifies $57\%$ of ground-truth LNs, highlighting the value of anatomical priors and high-quality annotations for challenging medical segmentation tasks, with code available at the provided repository.

Abstract

The evaluation of lymph node metastases plays a crucial role in achieving precise cancer staging, influencing subsequent decisions regarding treatment options. Lymph node detection poses challenges due to the presence of unclear boundaries and the diverse range of sizes and morphological characteristics, making it a resource-intensive process. As part of the LNQ 2023 MICCAI challenge, we propose the use of anatomical priors as a tool to address the challenges that persist in mediastinal lymph node segmentation in combination with the partial annotation of the challenge training data. The model ensemble using all suggested modifications yields a Dice score of 0.6033 and segments 57% of the ground truth lymph nodes, compared to 27% when training on CT only. Segmentation accuracy is improved significantly by incorporating a probabilistic lymph node atlas in loss weighting and post-processing. The largest performance gains are achieved by oversampling fully annotated data to account for the partial annotation of the challenge training data, as well as adding additional data augmentation to address the high heterogeneity of the CT images and lymph node appearance. Our code is available at https://github.com/MICAI-IMI-UzL/LNQ2023.

Figures (6)

• Figure 1: Visualization of the training pipeline.
• Figure 2: Distance maps overlaid over CT for example patients in axial, sagittal, and coronal view. Here, the contrasts of the distance map are set in a way that the contour of the same distance value for both patients is visualized in red. The reference point from which distances are measured is marked with a red cross. The distances are measured in the coordinate system of the atlas patient, thus, the contours show a deformed circle.
• Figure 3: Probabilistic lymph node atlases overlaid over CT for example patients in axial, sagittal, and coronal view.
• Figure 4: Learning curves for a semi-supervised approach. The blue and red curve show the train and the validation loss respectively. The green line is a moving average of the Dice metric. At 2,000 epochs, that is half training time, semi-supervised training starts.
• Figure 5: Probabilities for lymph node occurrence created from (a) the train, (b) the validation, and (c) the test dataset. The probabilities range from 0 to 0.85, the according color map is displayed on the right.