Abstract
This paper introduces $FSS_{BiGRU}$, a few-shot learning (FSL) framework designed for 3D medical image segmentation that bridges the gap between 2D efficiency and 3D structural awareness. By integrating a bidirectional Gated Recurrent Unit (GRU) into a U-Net-style architecture, the model learns spatial consistency between adjacent slices, allowing it to accurately segment new organs using as few as one to five annotated examples. The approach achieves performance comparable to fully supervised models while significantly reducing the need for manual annotation.
Introduction
Deep learning has revolutionized medical image segmentation, but it typically relies on massive, manually annotated datasets. In the medical field, this is a major bottleneck: experts must spend hours labeling 3D volumes, and models trained on one organ often fail to generalize to another. While few-shot learning aims to solve this by learning from very few examples, existing FSL methods for medical imaging often process 3D volumes as independent 2D slices. This ignores the "z-axis" context—the structural relationship between adjacent slices—leading to segmented outputs that are jagged, inconsistent, or anatomically unrealistic.
Methods
The authors propose a model called $FSS_{BiGRU}$, which consists of three main components:
- Dual Encoders: The model uses two separate VGG16-based encoders. One encoder ($E_s$) processes the "support" data (the few labeled examples provided), while the other ($E_q$) processes the "query" image (the new scan to be segmented).
- Bidirectional GRU (BiGRU) Bridge: This is the core innovation. Instead of passing features directly from the encoder to the decoder, the features from adjacent slices are fed into a BiGRU. This allows the model to "remember" information from previous and subsequent slices, ensuring that the predicted segmentation is spatially smooth and consistent across the 3D volume.
- Transfer Learning-based Adaptation: In "multi-shot" scenarios, the authors introduce a fine-tuning step. Before final inference, the model is briefly re-trained on the provided support samples to adapt its parameters to the specific characteristics of the target organ.
Key Insights and Ideas
- 3D Context via 2D Recurrence: Fully 3D convolutional networks are computationally expensive. By using a BiGRU to link 2D slices, the authors achieve 3D-like spatial awareness without the massive memory overhead of 3D kernels.
- Smoothness Matters: Standard few-shot models often produce "noisy" results at organ boundaries. The BiGRU acts as a temporal-spatial filter that enforces anatomical continuity.
Evaluation and Results
The model was evaluated across three major 3D CT datasets: BCV (Multi-atlas labeling Beyond the Cranial Vault), CTORG, and the Medical Segmentation Decathlon. The proposed $FSS_{BiGRU}$ model significantly outperformed existing state-of-the-art few-shot models:
- Accuracy: On the BCV dataset, the model achieved a mean Dice score of 0.866 with 5-shot learning, remarkably close to the 0.867 achieved by a fully supervised U-Net trained on the entire dataset.
- One-Shot Performance: Even with just a single labeled slice, the model outperformed competitors by a wide margin (e.g., achieving 0.816 for the spleen vs. 0.722 for the baseline).
- Generalization: The model maintained robust performance on external datasets (CTORG and Decathlon), proving it learns to relate support labels to query images rather than just memorizing training distributions.
Conclusion
$FSS_{BiGRU}$ provides a robust solution for 3D medical image segmentation in data-scarce environments. By leveraging bidirectional recurrence, it maintains anatomical consistency across slices. Future work will focus on developing a more universal model capable of handling different body parts and modalities, such as MRI and CT, simultaneously without task-specific adjustments.
Summary updated