AI Research Portal

This paper introduces FLoMo-Net, a novel modular framework for medical image segmentation (MIS) designed to address challenges like anatomical variability and ambiguous boundaries. The core idea is a task-adaptive Mixture of Experts (MoE) routing framework that incorporates frequency and uncertainty correction. FLoMo-Net achieves superior boundary-aware performance and faster inference with fewer parameters compared to existing 2D MIS models across four benchmarks.

The paper addresses several bottlenecks in existing medical image segmentation methods:

Suboptimal fusion of spatial and frequency-domain features, limiting the ability to capture rich structural and textural representations.
Difficulty in balancing fine local details with global context, especially with anatomical variability and ambiguous boundaries.
Limitations of CNNs in capturing long-range dependencies and the computational cost of attention mechanisms and Transformers.
Adaptability issues in state-space models due to fixed hidden states, limiting robustness to inter-patient variability.

A novel Local-Global Mixture of Experts (LoGo-MoE) encoder that adaptively routes features across specialized convolutional branches for efficient, anatomy-aware feature encoding.
A Dual-Attention Selective Aggregator (DASA) that jointly selects informative channels and spatial regions using frequency-guided modulation and adaptive spatial attention.
A Frequency-Aware Multi-Scale Refinement (FAMSR) module integrated at the bridge to refine edges and textures through explicit low/high-frequency decomposition.
A False Positive/False Negative Corrective Attention Module (FPFN-CAM) integrated across decoder stages to leverage uncertainty (derived from entropy and cosine dissimilarity) for residual correction and improved boundary delineation.
Demonstration of state-of-the-art (SOTA) performance with a lean parameter footprint and strong generalization across four MIS benchmarks.

FLoMo-Net is a U-shaped architecture comprising an encoder, a bridge, and a decoder. The pipeline involves:

Encoder (LoGo-MoE and DASA): The encoder uses LoGo-MoE for dynamic scale selection and DASA for spatial-frequency refinement. LoGo-MoE integrates multiple expert branches with varying kernel sizes and dilation rates, using an adaptive soft routing mechanism guided by global semantic content to dynamically weigh experts. It also incorporates a global context branch and an attention-driven residual gating mechanism to fuse local and global streams. DASA refines features by integrating Gated Frequency-aware Channel Attention (GFCA) and Adaptive Weighted Spatial Attention (AWSA), coupling spectral descriptors with spatial gating.
Bridge (FAMSR): The FAMSR module enhances features by capturing low/high-frequency textures through 2D-DCT branches and fusing them with dilated spatial context. It uses a spatial gate to sharpen boundary regions and refines features through frequency-enhanced and spatially gated components.
Decoder (FPFN-CAM): The decoder stages integrate LoGo-MoE (without downsampling), DASA, and the FPFN-CAM. FPFN-CAM corrects false positives (FPs) and false negatives (FNs) by modeling encoder-decoder feature discrepancies using entropy and cosine dissimilarity. It generates a correction mask to suppress spurious activations and recover missed regions.

Architecture:
- The FLoMo-Net is a U-shaped architecture inspired by UNet.
- It consists of an encoder, a bridge, and a decoder.
- Encoder: Composed of four stages, each using LoGo-MoE and deeper stages incorporating DASA. Features {E1, E2, E3, E4} are produced with halved resolution and doubled channels at each stage.
- LoGo-MoE: Integrates seven lightweight expert branches with varying kernel sizes and dilation rates. It uses adaptive soft routing based on global semantic content.
- DASA: Integrates GFCA and AWSA. GFCA uses 2D-DCT for spectral reasoning, and AWSA provides spatial localization. DASA is placed at encoder stages 3-4 and all decoder stages.
- FAMSR: Located at the bridge, it fuses frequency components (from 2D-DCT) with dilated spatial context.
- FPFN-CAM: Integrated into the decoder stages for FP/FN correction using entropy and cosine dissimilarity.
- Variants: FLoMo-NetB0 (2.06M parameters), FLoMo-NetB1 (7.99M parameters), FLoMo-NetB2 (31.42M parameters).
Key Equations:
- LoGo-MoE adaptive weight: π = Softmax(z), where z are routing logits.
- LoGo-MoE fused output: Ffused = G Flocal + (1 - G) F global, where G = σ(BN(Conv3×3(Flocal - Fglobal))).
- FPFN-CAM entropy map: H = - Σ Sc log(Sc + ε).
- FPFN-CAM cosine dissimilarity: C = 1 - (Ê' · D') / (||Ê'|| · ||D'||).
Loss/Objective: A hybrid loss combining Focal Tversky and BCE-Dice was employed.
Optimization:
- Optimizer: Nadam.
- Initial learning rate: 0.0004, reduced by a factor of 0.1 after 10 epochs.
- Training epochs: Up to 100 epochs with early stopping after 15 stagnant epochs.
- Batch size: 16.
Augmentation: Flipping, Gaussian noise, blurring, rotation, cropping, CLAHE, scaling, optical distortion, elastic transform, and random shadow.
Data Handling:
- Input resizing: 224 x 224 pixels.
- ACDC dataset: 3D-to-2D slice-level preprocessing.
- BUSI dataset: Excluded 133 non-lesion images for lesion segmentation.
Implementation Details:
- Framework: TensorFlow 2.10.0.
- Hardware: Two NVIDIA A6000 GPUs.
- Cross-validation: Five-fold cross-validation.

Datasets:
- Montgomery County chest X-ray set (MC)
- CVC-ClinicDB
- BUSI (Breast Ultrasound Images Dataset)
- ACDC (Automated Cardiac Diagnosis Challenge)
Metrics:
- Mean Dice Similarity Coefficient (mDSC)
- Mean Intersection over Union (mIoU)
- 95th percentile Hausdorff Distance (HD95)
Baselines: U-Net, DSCNet, EU2Net, UMNet, TransFuse, UNetR, Hiformer, PVT-EMCAD, TransUNet, MIST, UTANet, CAMS-Net, LightM-UNet.
Numeric Results:
- FLoMo-NetB2 achieved the highest mDSC and mIoU with the lowest HD95 across multiple datasets.
- For CVC-ClinicDB, FLoMo-NetB2 achieved mDSC 96.02%, mIoU 92.16%, and HD95 18.80 mm.
- For BUSI, FLoMo-NetB2 achieved mDSC 93.96%, mIoU 88.92%, and HD95 3.48 mm.
- For MC, FLoMo-NetB2 achieved mDSC 98.31%, mIoU 96.70%, and HD95 4.77 mm.
- For ACDC, FLoMo-NetB2 achieved mDSC 92.70%, mIoU 85.68%, and HD95 8.94 mm.
- Paired t-tests and Wilcoxon signed-rank tests indicated statistically significant gains for FLoMo-NetB2 on multiple datasets (p < 0.04).
- Ablation studies showed that removing any module reduced performance, with LoGo-MoE removal causing the largest decline.
- Soft routing consistently outperformed masked top-k routing strategies.
Runtime/Training Cost:
- FLoMo-NetB0: 2.06 million parameters, 6.82 Giga Floating Point Operations (G FLOPs), 3.36 ms inference time.
- FLoMo-NetB2: 31.42 million parameters, 104.05 G FLOPs, 14.49 ms inference time.
- FLoMo-NetB2 achieved SOTA performance with fewer parameters and faster inference than prior state-of-the-art 2D MIS models.

No explicit paper-reported limitations were found.

Method evidence: "FLoMo-Net achieves superior boundary-aware performance and faster inference with fewer parameters than prior state-of-the-art 2D MIS models." (Abstract)
Result evidence: "FLoMo-NetB2 consistently attains the highest mDSC and mIoU with the lowest HD95 (mm), reflecting superior accuracy and boundary fidelity; paired t-tests and Wilcoxon signed-rank tests on 5-fold results indicate statistically significant gains on multiple datasets (p < 0.04)." (Section 4. Experiments and Result Analysis)
Limitation evidence: No explicit paper-reported limitations were found.

FLoMo-Net: A Novel Task-Adaptive Mixture of Experts Routing Framework with Frequency and Uncertainty Correction for Medical Image Segmentation

Summary

Problem

Key Contributions

Method

Technical Details

Results

Weaknesses / Limitations

Evidence

Additional Figures