Core Concepts
A novel semi-supervised medical image segmentation framework, DiHC-Net, that leverages diagonal hierarchical consistency learning between multiple diversified sub-models to effectively utilize scarce labeled data and abundant unlabeled data.
Abstract
The paper proposes a novel semi-supervised medical image segmentation framework called DiHC-Net. The key aspects of the framework are:
-
Network Architecture:
- The network consists of three identical multi-scale V-Net sub-models with distinct sub-layers, such as upsampling and normalization, to increase intra-model diversity.
- The sub-models are trained using deep supervision on the labeled data, where the differences between the upsampled intermediate predictions and ground truth are minimized.
-
Diagonal Hierarchical Consistency Learning:
- To reduce inconsistencies between the sub-models' predictions, especially in challenging regions, the framework employs two consistency losses:
- Mutual Consistency Loss: Minimizes the difference between one sub-model's final prediction and the soft pseudo-labels from other sub-models.
- Diagonal Hierarchical Consistency Loss: Minimizes the difference between one sub-model's pseudo-labels and the intermediate and final representations of the other sub-models in a diagonal hierarchical fashion.
- The consistency losses are applied to both labeled and unlabeled data to leverage the abundant unlabeled data.
-
Experimental Validation:
- The proposed DiHC-Net framework is evaluated on two public medical image segmentation datasets: Left Atrium (LA) and Brain Tumor Segmentation (BraTS) 2019.
- DiHC-Net outperforms previous state-of-the-art semi-supervised methods across various performance metrics, demonstrating the effectiveness of the proposed approach.
The paper presents a simple yet effective semi-supervised medical image segmentation framework that leverages the diversity of sub-models and diagonal hierarchical consistency learning to achieve robust performance with limited labeled data.
Stats
The network is trained using a small set of labeled data (10% or 20%) and a large set of unlabeled data.
Quotes
"Accordingly, semi-supervised medical image segmentation (SSMIS) has undergone significant advancements."
"Motivated by recent advancements, we introduce a novel SSMIS framework under the assumption that a network, composed of diversified sub-models, can first fully learn from the scarce labelled data then collaborate by minimising disparities in predictions on uncertain regions yielded from both labelled and unlabelled data."