[Paper Review] GASNet: Weakly-supervised Framework for COVID-19 Lesion Segmentation

Zhanwei Xu, Yukun Cao+6 more|arXiv (Cornell University)|Oct 19, 2020

COVID-19 diagnosis using AI58 references22 citations

TL;DR

GASNet is a weakly-supervised deep learning framework for COVID-19 lung lesion segmentation that leverages volume-level diagnosis labels (positive/negative) and generative adversarial training to minimize reliance on costly voxel-level annotations. By jointly training a segmenter and generator to produce realistic 'healthy' CT volumes that fool a discriminator, GASNet achieves performance comparable to fully supervised models using only one voxel-level labeled sample.

ABSTRACT

Segmentation of infected areas in chest CT volumes is of great significance for further diagnosis and treatment of COVID-19 patients. Due to the complex shapes and varied appearances of lesions, a large number of voxel-level labeled samples are generally required to train a lesion segmentation network, which is a main bottleneck for developing deep learning based medical image segmentation algorithms. In this paper, we propose a weakly-supervised lesion segmentation framework by embedding the Generative Adversarial training process into the Segmentation Network, which is called GASNet. GASNet is optimized to segment the lesion areas of a COVID-19 CT by the segmenter, and to replace the abnormal appearance with a generated normal appearance by the generator, so that the restored CT volumes are indistinguishable from healthy CT volumes by the discriminator. GASNet is supervised by chest CT volumes of many healthy and COVID-19 subjects without voxel-level annotations. Experiments on three public databases show that when using as few as one voxel-level labeled sample, the performance of GASNet is comparable to fully-supervised segmentation algorithms trained on dozens of voxel-level labeled samples.

Motivation & Objective

• To address the high annotation cost of voxel-level segmentation in COVID-19 CT imaging, where expert labeling is time-consuming and inconsistent.
• To reduce dependency on large amounts of pixel/voxel-level annotated data for training deep segmentation networks.
• To develop a weakly-supervised framework that leverages readily available volume-level diagnosis labels (e.g., positive/negative) instead of expensive per-voxel annotations.
• To improve segmentation performance by integrating generative adversarial training into a segmentation network, enabling synthetic 'normal' CT volumes to be indistinguishable from real healthy scans.

Proposed method

• GASNet integrates a segmentation network (segmenter) and a generative network (generator) within a GAN framework, where both share the same input CT volume.
• The generator replaces lesion regions in a COVID-19 CT with synthesized 'normal' tissue, producing a synthetic healthy volume.
• The discriminator is trained to distinguish real healthy CT volumes from the synthetic ones, providing adversarial feedback to improve realism.
• The segmenter and generator are jointly optimized to minimize reconstruction loss, adversarial loss, and additional auxiliary losses (e.g., $σ_{IgToD}$, $σ_{IdToD}$, $σ_{MIL}$, $σ_{ps}$) for improved training stability and performance.
• The framework uses only volume-level labels (e.g., 'COVID-19 positive' or 'negative') for supervision, with minimal use of voxel-level annotations.
• Auxiliary losses such as $σ_{ps}$ provide direct voxel-level supervision to the segmenter, enhancing localization accuracy.

Experimental results

Research questions

• RQ1 Can a weakly-supervised segmentation framework achieve performance comparable to fully supervised models using only a single voxel-level labeled sample?
• RQ2 How effective is generative adversarial training in improving the realism of synthetic healthy CT volumes for lesion restoration?
• RQ3 To what extent do auxiliary loss functions improve training stability and segmentation accuracy in the absence of abundant pixel-level annotations?
• RQ4 Can the framework generalize across different datasets and maintain high performance with minimal human-annotated data?
• RQ5 How does the integration of volume-level diagnosis labels with adversarial training enhance segmentation performance in low-supervision regimes?

Key findings

• GASNet achieves a Dice score of 60.2% ± 23.4 on public Dataset-A using only one voxel-level labeled sample, outperforming fully supervised baselines trained on dozens of such samples.
• With one labeled sample, GASNet's Dice score (60.2%) is comparable to fully supervised 3D U-Net trained on 50 voxel-level labeled samples.
• The inclusion of auxiliary losses, especially $σ_{IgToD}$ and $σ_{IdToD}$, improved the Dice score by over 10 percentage points compared to the baseline without such constraints.
• The addition of $σ_{ps}$ further boosted the Dice score from 64.75% to 70.3%, demonstrating the value of direct voxel-level supervision in weakly-supervised learning.
• The fine-tuned version of GASNet ($\text{GASNet}_{\text{finetune}}$) achieved a Dice score of 58.9% ± 24.4% with 1 labeled sample, showing robustness and generalization across datasets.
• GASNet maintained high specificity (99.6% ± 0.2) and sensitivity (55.6% ± 28.3) on Dataset-A, indicating strong performance in distinguishing lesion from healthy tissue.