Technical Summary: HistoSeg++

Problem Statement

The segmentation of biomarkers in medical images is a foundational step for bioinformatics and biomedical applications. However, existing methods face significant challenges, particularly in capturing information across multiple scales and performing effective upsampling across diverse datasets. These limitations often lead to suboptimal generalization capabilities. While Nested-UNet architectures have shown promise in capturing multiscale contextual information, they often struggle with long-range dependencies (a limitation of CNNs) or precise local details (a limitation of pure Transformer approaches). Furthermore, biomarker segmentation is complicated by significant variations in size, appearance, and unclear boundaries.

Methodology

The authors propose HistoSeg++, a novel Nested-UNet architecture designed to enhance biomarker segmentation through hierarchical feature fusion and attention mechanisms. The architecture is structured as follows:

• Hierarchical Nested Design: The model employs an Outer-UNet scaffold that manages information flow through successive downsampling and upsampling stages. Embedded within this outer network are multiple Inner-UNet (I-UNet) modules. Each I-UNet module refines feature representations at specific resolutions by combining convolution blocks with attention gates.
• Attention Mechanisms: To focus on relevant features during upsampling, the model integrates two types of attention units:
  • Inner Attention: Located within the I-UNet blocks to filter noisy signals from skip connections and regulate gradient flow.
  • Outer Attention: Integrated into the main decoder path to prioritize key features.
• Feature Recalibration: The architecture utilizes Squeeze-and-Excitation (SE) modules at every stage to perform channel-wise feature recalibration, enhancing the model's ability to encode features.
• Multi-Scale Context Extraction:
  • Bridge Blocks: In deeper layers where spatial dimensions are small, dilated convolutions are used to learn dense semantic information.
  • Atrous Spatial Pyramid Pooling (ASPP): Modified with depthwise separable convolutions to extract high-semantic features with a larger field-of-view while maintaining spatial dimensions.
• Edge-Aware Loss: To improve boundary accuracy, the model employs a custom loss function that assigns greater importance to edge regions. This is achieved by identifying edges via Sobel filtering and amplifying errors at true object edges during the calculation of the binary cross-entropy loss.

Key Contributions

The paper outlines four primary contributions:

1. Hierarchical Architecture: A model with a hierarchical arrangement of encoder and decoder modules that efficiently integrates contextual features across multiple scales.
2. Dual Attention Units: The introduction of inner and outer attention units, enabling the network to focus on more relevant features specifically during the upsampling process.
3. Channel-Wise Recalibration: The integration of Squeeze-and-Excitation modules to further enhance segmentation quality.
4. Comprehensive Validation: Extensive experiments conducted on three public datasets demonstrating that these enhancements yield performance gains over existing Nested-UNet methods.

Experimental Results

The proposed method was evaluated on three publicly available benchmark datasets:

1. MoNuSeg: Multi-organ Nucleus Segmentation (H&E-stained histopathology).
2. DSB: 2018 Data Science Bowl (diverse cell images).
3. EM: Electron Microscopy (mitochondria in the CA1 hippocampus).

Quantitative Performance:

• MoNuSeg: HistoSeg++ achieved the highest IoU (71.44%) and Dice (83.30%), outperforming transformer-based methods like Swin-UNet and other Nested-UNet variants like UNet++. It also achieved the lowest HD95 (2.4399), indicating superior boundary accuracy.
• DSB: The model achieved an IoU of 85.50% and a Dice score of 91.81%, surpassing all baselines including TransUNet and nnU-Net.
• EM: HistoSeg++ attained an IoU of 86.54% and a Dice of 92.74%, demonstrating robust performance in mitochondrial segmentation where UNet struggled.

Ablation Studies:
Experiments on the MoNuSeg dataset confirmed that adding inner and outer attention units, ASPP, and SE modules progressively improved IoU and Dice scores while reducing HD95. The full model configuration yielded the best results.

Qualitative Analysis:
Visual comparisons indicated that while transformer-based methods tended to over-segment and introduce artifacts, and standard UNet variants struggled with fine details, HistoSeg++ produced reliable, coherent masks with minimal over-segmentation.

Significance and Claims

The paper claims that HistoSeg++ successfully addresses the shortcomings of existing Nested-UNet methods by effectively capturing multi-scale contextual information and enhancing upsampling through attention mechanisms. By integrating edge-aware loss and channel-wise recalibration, the model demonstrates superior generalization performance across diverse medical imaging scenarios. The authors position this work as a step toward more accurate biomarker segmentation, noting that with further refinement, the approach could support real-time processing and multi-class segmentation tasks. The code for the proposed method is made publicly available to facilitate reproducibility.