Abstract:Existing few-shot medical image segmentation (FSMIS) models fail to address a practical issue in medical imaging: the domain shift caused by different imaging techniques, which limits the applicability to current FSMIS tasks. To overcome this limitation, we focus on the cross-domain few-shot medical image segmentation (CD-FSMIS) task, aiming to develop a generalized model capable of adapting to a broader range of medical image segmentation scenarios with limited labeled data from the novel target domain. Inspired by the characteristics of frequency domain similarity across different domains, we propose a Frequency-aware Matching Network (FAMNet), which includes two key components: a Frequency-aware Matching (FAM) module and a Multi-Spectral Fusion (MSF) module. The FAM module tackles two problems during the meta-learning phase: 1) intra-domain variance caused by the inherent support-query bias, due to the different appearances of organs and lesions, and 2) inter-domain variance caused by different medical imaging techniques. Additionally, we design an MSF module to integrate the different frequency features decoupled by the FAM module, and further mitigate the impact of inter-domain variance on the model's segmentation performance. Combining these two modules, our FAMNet surpasses existing FSMIS models and Cross-domain Few-shot Semantic Segmentation models on three cross-domain datasets, achieving state-of-the-art performance in the CD-FSMIS task.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is the domain shift problem in cross - domain few - shot medical image segmentation (CD - FSMIS). Specifically, the existing few - shot medical image segmentation (FSMIS) models perform poorly when dealing with domain differences caused by different imaging techniques, which limits their wide applicability in practical applications. To solve this problem, the authors propose a new method named FAMNet (Frequency - aware Matching Network), aiming to train a general model that can adapt to a wider range of medical image segmentation scenarios with limited target - domain labeled data.
### Specific description of the problem
1. **Intra - domain Variance**:
- The manifestations of different organs and lesions in medical images vary greatly, resulting in support - query bias between support sets, thus affecting the quality of prototype representation.
2. **Inter - domain Variance**:
- There is a significant problem of low spatial - domain similarity between different imaging techniques (such as CT and MRI), especially between high - frequency and low - frequency bands. However, in the frequency domain, the similarity in the mid - frequency band is relatively high.
### Solution
To address the above challenges, the authors propose a new method, FAMNet, which includes two key modules:
1. **Frequency - aware Matching (FAM) Module**:
- The FAM module reduces the support - query bias by performing support - query matching on specific frequency bands and fuses foreground features, highlighting the collaborative parts.
- At the same time, the FAM module introduces frequency - domain information into the feature space, reducing the dependence on frequency bands with significant domain differences, enabling the model to focus more on domain - independent frequency bands.
2. **Multi - spectral Fusion (MSF) Module**:
- The MSF module fuses different frequency features decoupled by the FAM module and further extracts key information in domain - specific frequency bands, thereby reducing the impact of inter - domain variation on the model's segmentation performance.
### Summary
By combining these two modules, FAMNet achieves state - of - the - art performance on three cross - domain datasets, significantly improving the performance of cross - domain few - shot medical image segmentation tasks. This method not only solves the domain shift problem but also demonstrates strong generalization ability and excellent segmentation performance.
### Formula summary
- Formula for generating support foreground prototypes:
\[
p_{fg}^s=\frac{\sum_{u,v}F_s^{ini}(u,v)M_s(u,v)}{\sum_{u,v}M_s(u,v)}
\]
- Formula for predicting coarse query foreground masks:
\[
\tilde{M}_{Coarse}^q = 1.0-\sigma\left(\frac{d(F_q^{ini},p_{fg}^s)-\tau}{\alpha}\right)
\]
- Formula for frequency - band decomposition:
\[
F_s(i)=\rho^{-1}(F^{-1}(B_p(\phi_s,I(i))))
\]
- Attention - weighted formula (DAFBs):
\[
F''_s,0 = A'\odot F'_s
\]
- Reverse attention - weighted formula (DSFBs):
\[
F''_s,1=(1 - A')\odot F'_s
\]
The introduction of these formulas enables FAMNet to effectively handle intra - domain and inter - domain differences in the frequency domain, thereby improving the model's generalization ability and segmentation accuracy.