Uncertainty-Aware Temporal Self-Learning (UATS): Semi-Supervised Learning for Segmentation of Prostate Zones and Beyond

Anneke Meyer, Suhita Ghosh, Daniel Schindele, Martin Schostak, Sebastian Stober, Christian Hansen, Marko Rak

Introduction

Deep learning has revolutionized medical imaging, but its progress is often limited by the "data hunger" of supervised models. Training a model to segment the prostate into its four distinct anatomical zones—the transition zone (TZ), peripheral zone (PZ), distal prostatic urethra (DPU), and anterior fibromuscular stroma (AFS)—requires massive amounts of pixel-perfect labels. In the clinical world, these labels must be created by highly trained radiologists, making them expensive and scarce.

While most existing AI models only tackle the two largest zones (TZ and PZ), this research introduces Uncertainty-Aware Temporal Self-Learning (UATS), a framework designed for fine-grained segmentation of all four zones by effectively leveraging the mountains of unlabeled medical images sitting in hospital archives.

The UATS Framework

The authors propose a two-stage training strategy to maximize the utility of limited labeled datasets:

  1. Stage I (Warm-up): A standard 3D U-Net is trained on a small set of labeled images until it converges, establishing a baseline understanding of the anatomy.
  2. Stage II (Semi-Supervised): The model continues training using both labeled and unlabeled images. This stage introduces two critical mechanisms:
  3. Temporal Ensembling: Instead of relying on the model's prediction from a single training step, UATS maintains a weighted moving average of predictions over several epochs. This "ensemble" acts as a more stable teacher, filtering out the noise of early training stages.
  4. Uncertainty Guidance: The model calculates a confidence score for its predictions. It only generates "pseudo-labels" for the pixels it is most certain about, preventing the model from reinforcing its own mistakes (confirmation bias).

To further improve accuracy, the authors use a Continuous Dice Coefficient (cDC) loss function. This specialized loss handles probabilistic (fuzzy) predictions rather than just hard binary values, preserving more information during the self-learning process.

Axial and sagittal views of the four-zone prostate segmentation produced by the UATS framework.

Experimental Results

The researchers validated UATS across three distinct tasks to prove its versatility: prostate segmentation (ProstateX), hippocampus segmentation (Medical Decathlon), and skin lesion segmentation (ISIC 2018).

Key Findings:

  • Prostate Performance: UATS achieved Dice Coefficients (a measure of overlap where 100% is perfect) of 87.3% for the TZ and 78.9% for the PZ.
  • The Minority Classes: The biggest wins were in the difficult, small structures. The AFS zone saw a jump to 50.6% DC, and the DPU reached 75.3% DC, significantly outperforming the supervised-only baseline.
  • Human-Level Performance: For the first time in fine-grained prostate segmentation, the AI's performance reached the "inter-rater" level—meaning the AI's disagreement with a doctor was no greater than the disagreement between two different doctors.
  • Resilience: When tested with Gaussian noise, the UATS model proved much more resilient than standard supervised models, suggesting better generalization to real-world clinical data.

Conclusion and Future Work

UATS demonstrates that semi-supervised learning is most valuable when labeled data is at its scarcest, making it an ideal candidate for rare diseases or new imaging modalities. While performance on the Peripheral Zone (PZ) remains slightly below human levels due to "partial volume effects" (blurring at slice edges), the framework provides a robust path forward for scaling medical AI. Future research may explore predicting beforehand if semi-supervised learning will benefit a specific task by measuring initial model uncertainty across unlabeled sets.

Summary updated

Abstract

Various convolutional neural network (CNN) based concepts have been introduced for the prostate's automatic segmentation and its coarse subdivision into transition zone (TZ) and peripheral zone (PZ). However, when targeting a fine-grained segmentation of TZ, PZ, distal prostatic urethra (DPU) and the anterior fibromuscular stroma (AFS), the task becomes more challenging and has not yet been solved at the level of human performance. One reason might be the insufficient amount of labeled data for supervised training. Therefore, we propose to apply a semi-supervised learning (SSL) technique named uncertainty-aware temporal self-learning (UATS) to overcome the expensive and time-consuming manual ground truth labeling. We combine the SSL techniques temporal ensembling and uncertainty-guided self-learning to benefit from unlabeled images, which are often readily available. Our method significantly outperforms the supervised baseline and obtained a Dice coefficient (DC) of up to 78.9% , 87.3%, 75.3%, 50.6% for TZ, PZ, DPU and AFS, respectively. The obtained results are in the range of human inter-rater performance for all structures. Moreover, we investigate the method's robustness against noise and demonstrate the generalization capability for varying ratios of labeled data and on other challenging tasks, namely the hippocampus and skin lesion segmentation. UATS achieved superiority segmentation quality compared to the supervised baseline, particularly for minimal amounts of labeled data.
BibTeX
@inproceedings{meyer2021uncertainty,
  author = {Anneke Meyer and Suhita Ghosh and Daniel Schindele and Martin Schostak and Sebastian Stober and Christian Hansen and Marko Rak},
  title = {{Uncertainty-Aware Temporal Self-Learning (UATS): Semi-Supervised Learning for Segmentation of Prostate Zones and Beyond}},
  booktitle = {Artif. Intell. Medicine},
  year = {2021},
  doi = {10.1016/j.artmed.2021.102073},
}
Figure