Source-linked AI summary
RAUNet: Residual Attention U-Net for Semantic Segmentation of Cataract Surgical Instruments
Zhen-Liang Ni, Gui-Bin Bian, Xiao-Hu Zhou, Zeng-Guang Hou, Xiao-Liang Xie, Chen Wang, Yan-Jie Zhou, Rui-Qi Li, Zhen Li
TL;DR
Cataract surgical-instrument segmentation remains difficult because of specular reflection and class imbalance. RAUNet combines an augmented attention module, hybrid loss, and the Cata7 dataset, achieving state-of-the-art performance with 97.71% mean Dice and 95.62% mean IOU.
Problem
Accurate cataract surgical-instrument segmentation is challenging because of specular reflection and class imbalance, despite its relevance to robot-assisted surgery.
Method
RAUNet uses an augmented attention module to emphasize target features, a hybrid cross-entropy and logarithmic Dice loss, and the Cata7 semantic-segmentation dataset.
Results
97.71% mean Dice and 95.62% mean IOU give RAUNet state-of-the-art performance on Cata7.
Takeaways & Limitations
The attention module improves segmentation accuracy while adding very few parameters, and the hybrid loss addresses class imbalance.
Abstract
from arXiv · showhide
Semantic segmentation of surgical instruments plays a crucial role in robot-assisted surgery. However, accurate segmentation of cataract surgical instruments is still a challenge due to specular reflection and class imbalance issues. In this paper, an attention-guided network is proposed to segment the cataract surgical instrument. A new attention module is designed to learn discriminative features and address the specular reflection issue. It captures global context and encodes semantic dependencies to emphasize key semantic features, boosting the feature representation. This attention module has very few parameters, which helps to save memory. Thus, it can be flexibly plugged into other networks. Besides, a hybrid loss is introduced to train our network for addressing the class imbalance issue, which merges cross entropy and logarithms of Dice loss. A new dataset named Cata7 is constructed to evaluate our network. To the best of our knowledge, this is the first cataract surgical instrument dataset for semantic segmentation. Based on this dataset, RAUNet achieves state-of-the-art performance 97.71% mean Dice and 95.62% mean IOU.
1 Introduction
Cataract surgical-instrument segmentation supports robot-assisted surgery but remains underexplored and difficult because strong lighting causes specular reflection and small instruments create severe class imbalance. RAUNet addresses these challenges with attention-guided feature representation, a hybrid loss, and the Cata7 dataset.
- Research motivation: Cataract surgery has few dedicated segmentation studies despite its importance for computer-assisted surgery and robot control.The paper presents this work as the first study to segment and classify cataract surgical instruments.
- Research motivation: Specular reflection and the dominance of background pixels make cataract surgical instruments difficult to identify and segment.Strong lighting changes instrument appearance, while small instruments occupy limited image regions and create class imbalance.
- Proposed approach: RAUNet introduces an augmented attention module to improve feature representation and address the specular-reflection challenge.The module is designed to fuse multi-level features while using very few parameters.
- Proposed approach: The proposed framework combines a hybrid cross-entropy and logarithmic Dice loss with the Cata7 dataset for cataract-instrument segmentation.Cata7 is presented as the first cataract surgical-instrument dataset for semantic segmentation.
2 Residual Attention U-Net
RAUNet combines a ResNet34 encoder-decoder with an augmented attention module that fuses semantic and location information, plus a hybrid loss targeting class imbalance. The design uses global context and lightweight operations to emphasize instrument features while preserving position details.
- Architecture: RAUNet uses a ResNet34 encoder-decoder architecture to produce high-resolution instrument masks while reducing model size and improving inference speed.The decoder upsamples features and uses attention-guided fusion rather than naive skip concatenation.
- Augmented Attention Module: The augmented attention module captures high-level semantic information and low-level global context to emphasize target features and suppress interfering background information.High-level features guide the selection of important location details in low-level maps, while channel dependencies encode semantic relationships.
- Augmented Attention Module: Global average pooling and 1×1 convolutions generate an attentive vector that encodes semantic dependencies, and element-wise multiplication calibrates low-level features before addition with high-level features.Addition replaces concatenation to reduce convolutional parameters and computational cost.
- Loss Function: The hybrid CEL-Dice loss combines cross entropy with logarithmic Dice loss to retain cross-entropy stability while addressing foreground-background class imbalance.The logarithm increases the penalty when Dice similarity is small, making poor predictions more strongly penalized.
- Cata7 Dataset: Cata7 contains 2,500 images split into training and test sets, with up to two surgical instruments per image and ten instrument categories.The dataset is evaluated using images from five training video sequences and two test video sequences.
3 Experiments
Experiments evaluate RAUNet on the Cata7 dataset through dataset construction, attention-module ablations, comparisons with other methods, class-wise accuracy, visualizations, and loss-function analysis.
- Datasets: Cata7 contains 2,500 images from seven complete cataract-surgery videos, with five sequences for training and two for testing.Images are labeled with precise instrument edges and types; ten instrument categories are represented.
- Ablation for augmented attention module: 2.59% mean Dice and 4.31% mean IOU gains result from adding AAM to the base network, while increasing parameters by only 0.26M.AAM also outperforms GAU while adding fewer parameters.
- Ablation for augmented attention module: RAUNet visualizations show complete instrument segmentation matching the ground truth, whereas the base network exhibits classification errors and incomplete masks.The visual comparisons are presented in Fig. 4.
- Comparison with state-of-the-art: 97.71% mean Dice and 95.62% mean IOU make RAUNet the reported state-of-the-art on Cata7, outperforming U-Net, TernausNet, and LinkNet.The comparison is reported in Table 3.
- Comparison with state-of-the-art: Lens-hook and primary-incision-knife pixel accuracy reaches 90.23% and 100%, respectively, despite misclassification difficulties for competing methods.The reported errors are associated with thin instruments, limited samples, and insufficient high-level semantic information.
- Verify the performance of CEL-Dice: CEL-Dice produces higher mean Dice and mean IOU than Dice loss and cross entropy across training epochs.The loss combines cross entropy with logarithmic Dice loss for class imbalance.
4 Conclusion
RAUNet combines an augmented attention module with Cross Entropy Log Dice loss for cataract surgical-instrument segmentation, achieving state-of-the-art performance on Cata7.
- RAUNet achieves state-of-the-art performance on the Cata7 dataset for surgical-instrument segmentation.
- The augmented attention module improves feature representation and segmentation accuracy while adding very few parameters.
- Cross Entropy Log Dice is introduced to address the class imbalance issue.