Source-linked AI summary
CaraNet: Context Axial Reverse Attention Network for Segmentation of Small Medical Objects
Ange Lou, Shuyue Guan, Hanseok Ko, Murray Loew
TL;DR
Medical segmentation models have rarely focused on small objects, despite the importance of accurate segmentation across varied object sizes and imaging modalities. CaraNet combines axial reverse attention with a channel-wise feature pyramid and is evaluated across polyp and brain-tumor datasets. The paper reports top-rank mean Dice performance and advantages over prior models for small objects, while identifying limitations in upsampling, pretraining, and sliced MRI data.
Problem
Existing medical segmentation networks seldom consider small objects, although object size varies widely and small-object accuracy matters for disease diagnosis and treatment.
Method
CaraNet combines a pretrained backbone, partial decoder, channel-wise feature pyramid, and axial reverse attention to segment small medical objects.
Results
CaraNet outperforms compared models overall and on small objects across polyp and brain-tumor segmentation datasets, including at least 2% higher overall mean Dice than state-of-the-art approaches.
Takeaways & Limitations
CaraNet shows a distinct advantage for segmenting small single, multi-object, and extremely small medical targets within the evaluated datasets.
Takeaways & Limitations
The authors identify coarse boundaries from bilinear upsampling, ImageNet pretraining unlike medical images, and spatial-information loss from sliced brain MRI data as limitations.
Abstract
from arXiv · showhide
Segmenting medical images accurately and reliably is important for disease diagnosis and treatment. It is a challenging task because of the wide variety of objects' sizes, shapes, and scanning modalities. Recently, many convolutional neural networks (CNN) have been designed for segmentation tasks and achieved great success. Few studies, however, have fully considered the sizes of objects, and thus most demonstrate poor performance for small objects segmentation. This can have a significant impact on the early detection of diseases. This paper proposes a Context Axial Reserve Attention Network (CaraNet) to improve the segmentation performance on small objects compared with several recent state-of-the-art models. We test our CaraNet on brain tumor (BraTS 2018) and polyp (Kvasir-SEG, CVC-ColonDB, CVC-ClinicDB, CVC-300, and ETIS-LaribPolypDB) segmentation datasets. Our CaraNet achieves the top-rank mean Dice segmentation accuracy, and results show a distinct advantage of CaraNet in the segmentation of small medical objects.
1. INTRODUCTION
Medical image segmentation has advanced across imaging modalities, but existing networks rarely address the specific difficulty of segmenting small medical objects. CaraNet is proposed to target this gap and is reported to improve polyp segmentation overall and for small objects.
- CNN-based and attention-based networks have achieved strong medical image segmentation performance across modalities including X-ray, MRI, PET, CT, and visible-light imaging.
- Attention-based models such as MedT, TransUNet, and CoTr have improved segmentation compared with purely convolutional networks such as U-Net and FCN.
- Existing neural networks seldom consider small-object segmentation, particularly in medical imaging.
- CaraNet is introduced as an attention-based deep neural network for segmenting small medical objects.
- The paper reports that CaraNet outperforms most current models, including PraNet, overall and on small polyp objects.
2. METHOD
CaraNet combines a pretrained backbone, partial decoder, channel-wise feature pyramid, and axial reverse attention to aggregate semantic, multi-scale, and localization information. Its training uses weighted IoU/BCE losses with deep supervision, while evaluation analyzes mean Dice by object-size ratio.
- Architecture: CaraNet uses a pretrained Res2Net backbone, parallel partial decoder, CFP module, and A-RA module.The architecture overview identifies these four principal components.
- Partial decoder: The parallel partial decoder aggregates high-level Res2Net features {f3, f4, f5} to generate a global map S_g.High-level features are selected because low-level features contribute less performance while costing more computation at larger spatial resolutions.
- Channel-wise feature pyramid: The CFP module uses K=4 channels with dilation rates {r1, r2, r3, r4}={1,2,4,8}, hierarchical feature fusion, and regular or asymmetric convolutions.Each channel has dimension M/4, and skip connections connect convolutional-module outputs.
- Axial reverse attention module: A-RA applies axial attention to CFP multi-scale features and reverse attention to side-output salience information.Axial attention factorizes 2D attention into height- and width-axis 1D operations and replaces softmax with sigmoid.
- Deep supervision: Training combines weighted IoU and weighted BCE losses with deep supervision on three side-outputs and the global map after up-sampling to ground-truth size.
- Small object segmentation analysis: Object size is evaluated as the ratio m/N of object pixels m to total image pixels N, with primary focus on ratios below 5%.Mean-Dice coefficients are plotted against increasingly sorted size ratios and smoothed using interval averages.
3. EXPERIMENT
CaraNet was evaluated on five polyp datasets and BraTS 2018, with comparisons against established segmentation models and analyses organized by object size. Results indicate stronger performance for small and extremely small objects.
- Datasets: CaraNet was tested on five polyp datasets and the BraTS 2018 brain tumor dataset.The polyp datasets were ETIS, CVC-ClinicDB, CVC-ColonDB, EndoScene, and Kvasir.
- Comparison models: The evaluation compared CaraNet with six medical image segmentation models, including U-Net, UNet++, ResUNet-mod, ResUNet++, SFA, and PraNet.
- Metrics: The study reported mean Dice, mean IoU, weighted Dice, MAE, enhanced alignment, and structural similarity metrics.Weighted Dice addresses equal importance in Dice, MAE measures pixel-to-pixel accuracy, enhanced alignment measures pixel- and global-level similarity, and structural similarity compares prediction and ground-truth structure.
- Polyp results: CaraNet outperformed the compared models overall and on samples with small polyps across the five polyp datasets.Figure 7 compares CaraNet and PraNet for objects with proportions ≤5%.
- Brain tumor results: 3231 BraTS 2018 testing images contained tumor proportions ranging from 0.01%–4.91%, enabling analysis of extremely small tumors.CaraNet was compared with PraNet because PraNet had performance closest to CaraNet among the compared models.
- Brain tumor results: In the brain tumor size analysis, CaraNet performed better than PraNet, especially for tumors in the 0.01%–0.1% range.Figure 8 encodes the difference between CaraNet and PraNet mean Dice values; red regions indicate CaraNet is higher.
4. DISCUSSION
The discussion attributes CaraNet’s reported gains to its architecture and highlights limitations involving upsampling, pretraining, and loss of spatial information from sliced MRI data. It also questions the basis of the 5% small-object threshold.
- Contribution: CaraNet combines Axial Reverse Attention and Channel-wise Feature Pyramid modules to improve small medical object segmentation.
- Reported performance: CaraNet outperformed prominent models overall across six measurement metrics and produced precise predictions for challenging cases compared with PraNet.
- Limitations: Bilinear interpolation for upsampling can lose useful information and produce coarse boundaries; deconvolution is proposed as a possible improvement.
- Limitations: ImageNet pretraining uses natural images that differ from medical images, while sliced brain MRI data lose spatial information between voxels.The paper states that voxel-level spatial-information loss may influence small-tumor detection accuracy.
- Evaluation scope: The study defines small objects as having size ratios below 5% but does not provide a specific rationale or clear general definition for that threshold.
5. CONCLUSION
CaraNet is presented as a neural network for small medical object segmentation. The conclusion reports gains over state-of-the-art approaches overall and for small objects across the evaluated datasets.
- Overall conclusion: CaraNet outperformed all state-of-the-art approaches by at least 2% in mean Dice accuracy overall.
- Polyp conclusion: On ETIS, CaraNet reached 74.7% mean Dice accuracy, about 12% higher than PraNet.ETIS contains many small polyps.
- Brain tumor conclusion: On BraTS 2018, CaraNet achieved 3% higher performance than PraNet for extremely small object segmentation.
- Size-based conclusion: CaraNet outperformed PraNet for small objects in all six datasets used.