Fuzzy Attention-based Border Rendering Network for Lung Organ Segmentation

TL;DR

This work addresses automatic lung segmentation on CT where undefined voxel ranges, class imbalance, and slender airway/artery trees cause false negatives/positives and discontinuities. It introduces FABR, a two-module approach combining a fuzzy attention-based transformer-like backbone with a Global-Local Cube-tree Fusion module that concentrates on border vulnerable points and fuses global and local context. Across BAS, PARSE22, AeroPath, and Lung Fibrosis datasets, FABR delivers state-of-the-art performance, notably improving border accuracy and reducing leakage while maintaining high IoU and Dice scores. By explicitly modeling border regions and uncertainty in feature representations, FABR offers robust, border-aware 3D lung segmentation suitable for diverse clinical scenarios.

Abstract

Automatic lung organ segmentation on CT images is crucial for lung disease diagnosis. However, the unlimited voxel values and class imbalance of lung organs can lead to false-negative/positive and leakage issues in advanced methods. Additionally, some slender lung organs are easily lost during the recycled down/up-sample procedure, e.g., bronchioles & arterioles, causing severe discontinuity issue. Inspired by these, this paper introduces an effective lung organ segmentation method called Fuzzy Attention-based Border Rendering (FABR) network. Since fuzzy logic can handle the uncertainty in feature extraction, hence the fusion of deep networks and fuzzy sets should be a viable solution for better performance. Meanwhile, unlike prior top-tier methods that operate on all regular dense points, our FABR depicts lung organ regions as cube-trees, focusing only on recycle-sampled border vulnerable points, rendering the severely discontinuous, false-negative/positive organ regions with a novel Global-Local Cube-tree Fusion (GLCF) module. All experimental results, on four challenging datasets of airway & artery, demonstrate that our method can achieve the favorable performance significantly.

Figures (9)

• Figure 1: The elaboration of Border Vulnerable Points (BVP) caused by recycled down-sample and up-sample in the encoder-decoder backbone. Downsampling (c) gets (d), upsampling (d) gets (e), then (f) is the absolute difference of (c) & (e). In the test phase, (c) is binarized coarse prediction.
• Figure 2: The overview of our method FABR. FGLC: fine grain local context; CTCF: cube-tree centroid feature; CGLC: coarse grain local context; PLGC: projected learnable global context. BVP detector is shown in Fig. \ref{['fig:bvp']}. Noting the matched relationship between top-right boxes' and bottom-right bars' colors.
• Figure 3: Our FA-based transformer-like backbone design and coarse mask generation. FA: fuzzy attention module detailed in subsection \ref{['sec:fa']}. DW: depth-wise convolution.
• Figure 4: The details of (a) our efficient fuzzy attention module and (b) fuzzy attention gate (FAG) in the subfigure (a). Zooming in for a better view.
• Figure 5: Qualitative airway segmentation on BAS/Lung fibrosis datasets. GT: ground truth. Red color: true positive. Green color: false positive. Blue color: false negative.