Weakly Supervised Lymph Nodes Segmentation Based on Partial Instance Annotations with Pre-trained Dual-branch Network and Pseudo Label Learning

TL;DR

This work tackles automatic lymph node segmentation under partial annotations, a scenario that substantially reduces labeling cost yet challenges model recall. It introduces a pre-trained Dual-Branch network with Dynamically Mixed Pseudo labels (DBDMP) that combines self-supervised pretraining (Model Genesis) with online pseudo-label learning, using two decoders to generate robust soft pseudo labels for unannotated nodes. The method employs a carefully designed loss suite, including $\\mathcal{L}_{SCE}$, $\mathcal{L}_{PCE}$, $\mathcal{L}_{Tversky}$, and a consensus-aware $\mathcal{L}_{KLCE}$-based weighting, along with a ramp-up for pseudo-label supervision, yielding large gains over partial-annotation baselines. On the LNQ dataset, the approach achieves a Dice score of up to $57.36\%$ on the test set and reduces average symmetric surface distance to $9.35$ mm, highlighting its potential for clinically useful segmentation while reducing annotation burden.

Abstract

Assessing the presence of potentially malignant lymph nodes aids in estimating cancer progression, and identifying surrounding benign lymph nodes can assist in determining potential metastatic pathways for cancer. For quantitative analysis, automatic segmentation of lymph nodes is crucial. However, due to the labor-intensive and time-consuming manual annotation process required for a large number of lymph nodes, it is more practical to annotate only a subset of the lymph node instances to reduce annotation costs. In this study, we propose a pre-trained Dual-Branch network with Dynamically Mixed Pseudo label (DBDMP) to learn from partial instance annotations for lymph nodes segmentation. To obtain reliable pseudo labels for lymph nodes that are not annotated, we employ a dual-decoder network to generate different outputs that are then dynamically mixed. We integrate the original weak partial annotations with the mixed pseudo labels to supervise the network. To further leverage the extensive amount of unannotated voxels, we apply a self-supervised pre-training strategy to enhance the model's feature extraction capability. Experiments on the mediastinal Lymph Node Quantification (LNQ) dataset demonstrate that our method, compared to directly learning from partial instance annotations, significantly improves the Dice Similarity Coefficient (DSC) from 11.04% to 54.10% and reduces the Average Symmetric Surface Distance (ASSD) from 20.83 $mm$ to 8.72 $mm$. The code is available at https://github.com/WltyBY/LNQ2023_training_code.git

Figures (4)

• Figure 1: An overview of the proposed DBDMP which utilizes a dual-branch network with one shared encoder and two decoders. (a) In the self-supervised learning stage, Model Genesis is employed for pre-training. (b) In the downstream learning stage, a mixture of the outputs from the two decoders is combined with the original partial annotation to obtain a pseudo label. We also use a consensus-aware loss $\mathcal{L}_{KLCE}$ to avoid over-fitting noise in the pseudo labels.
• Figure 2: The transformations used to generate the input for self-supervised training: I. Non-linear transformation, II. Local pixel shuffling, III. In-painting, IV. Out-painting. RCT (Randomly Composed Transformation) means that the basic transformations are composed, each with a probability to be used. Note that in-painting and out-painting are not performed together each time.
• Figure 3: Sensitivity analysis of hyper-parameters $\tau$, $\gamma$, $\alpha$ and $\lambda$, respectively.
• Figure 4: Visualization of segmentation results of the ablation study in Table \ref{['tab:Ablation Study']}.