[Paper Review] More unlabelled data or label more data? A study on semi-supervised laparoscopic image segmentation

Yunguan Fu, Maria Robu+7 more|arXiv (Cornell University)|Aug 20, 2019

Advanced Neural Network Applications8 references4 citations

TL;DR

This study evaluates semi-supervised learning for laparoscopic liver segmentation using the mean teacher framework, demonstrating that adding unlabelled data can achieve segmentation accuracy comparable to labelling more data. It quantitatively shows that semi-supervised training with unlabelled data improves performance beyond supervised learning, even without additional labels, offering a practical trade-off for clinical data efficiency.

ABSTRACT

Improving a semi-supervised image segmentation task has the option of adding more unlabelled images, labelling the unlabelled images or combining both, as neither image acquisition nor expert labelling can be considered trivial in most clinical applications. With a laparoscopic liver image segmentation application, we investigate the performance impact by altering the quantities of labelled and unlabelled training data, using a semi-supervised segmentation algorithm based on the mean teacher learning paradigm. We first report a significantly higher segmentation accuracy, compared with supervised learning. Interestingly, this comparison reveals that the training strategy adopted in the semi-supervised algorithm is also responsible for this observed improvement, in addition to the added unlabelled data. We then compare different combinations of labelled and unlabelled data set sizes for training semi-supervised segmentation networks, to provide a quantitative example of the practically useful trade-off between the two data planning strategies in this surgical guidance application.

Motivation & Objective

• To evaluate the impact of increasing labelled versus unlabelled data on semi-supervised laparoscopic liver segmentation performance.
• To isolate the contribution of the mean teacher training strategy from the effect of added unlabelled data.
• To provide a quantitative basis for data planning decisions in clinical semi-supervised learning applications.
• To assess whether unlabelled data can achieve performance equivalent to labelling more data in a real surgical imaging context.

Proposed method

• Uses a U-Net-based segmentation network with multi-scale input pyramids and $L_2$-regularized soft Dice loss for binary segmentation.
• Employs the mean teacher framework with two identical networks: a student with trainable weights and a teacher with momentum-updated weights.
• Applies a combined loss: supervised loss $L_l$ on labelled data and unsupervised consistency loss $L_u$ between student and teacher predictions on unlabelled data, balanced by hyperparameter $\lambda$.
• Trains the student network using backpropagation with the total loss $L_s = L_l + \lambda L_u$, where $L_u$ penalizes prediction disagreement between perturbed views of the same unlabelled input.
• Uses data augmentation via random noise injection ($\eta$) to generate diverse views for consistency regularization.
• Evaluates performance across varying combinations of labelled and unlabelled data sizes using Dice score and Hausdorff distance on independent test sets.

Experimental results

Research questions

• RQ1 Does semi-supervised learning with the mean teacher framework improve laparoscopic liver segmentation accuracy compared to supervised learning?
• RQ2 To what extent does the training strategy of the mean teacher method contribute to performance gains beyond the addition of unlabelled data?
• RQ3 What is the quantitative trade-off between acquiring more labelled data versus more unlabelled data in terms of segmentation performance?
• RQ4 Can a semi-supervised model trained on a large amount of unlabelled data match or exceed the performance of a supervised model trained on more labelled data?

Key findings

• The mean teacher model achieved a median Dice score of 0.9646, significantly higher than the supervised baseline’s 0.9594 (p = 9.27e-28).
• The mean teacher model reduced median Hausdorff distance from 148.23 to 81.49, indicating improved boundary localization (p = 4.40e-07).
• Even with zero unlabelled data, the mean teacher model outperformed the supervised model, demonstrating that the training strategy itself improves performance.
• Using 100% unlabelled data and 50% labelled data, the mean teacher achieved a Dice score of 0.9611, exceeding the supervised model’s 0.9594 with 100% labelled data.
• Performance did not always improve monotonically with more unlabelled data, suggesting potential interference from intra-procedural data correlation or suboptimal hyperparameter tuning.
• The results provide the first quantitative evidence on real patient data that unlabelled data can substitute for labelled data in clinical semi-supervised segmentation tasks.