Abstract
Accurate segmentation of kidney tumors in CT images is a challenging task due to irregular motion, similar appearances to surrounding tissue, and diverse tumor shapes. This paper presents Crossbar-Net, a novel automatic segmentation method. The approach innovatively utilizes orthogonal non-squared patches, termed "crossbar patches," to capture both detailed local information and global contextual appearance in horizontal and vertical directions. Furthermore, a cascaded training strategy is employed to integrate results from two sub-models, ensuring consistency and improved accuracy. Experimental results on a dataset of 3,500 images from 94 patients demonstrate that Crossbar-Net achieves superior performance compared to state-of-the-art methods in terms of Dice ratio, Hausdorff distance, and other key metrics.
Introduction
Kidney tumor segmentation is a critical step in cancer diagnosis and treatment planning. However, it remains a difficult problem for automated systems. Tumors often exhibit low contrast with adjacent healthy tissue and vary significantly in size and morphology. Traditional Convolutional Neural Networks (CNNs) typically rely on square patches for training. While small square patches capture fine details, they lack the global context necessary to locate the tumor within the organ. Conversely, large square patches introduce significant noise and irrelevant information from distant regions.
To overcome these limitations, the authors propose Crossbar-Net. This framework leverages the inherent symmetry and axial structure of kidney tumors by using elongated rectangular patches that span the horizontal and vertical axes of the image.
Methodology: Crossbar Patches and Cascaded Training
The core innovation of Crossbar-Net lies in its sampling strategy and its collaborative training architecture.
Crossbar Patch Sampling
Instead of standard square regions, the model samples two orthogonal, non-squared patches centered on the same pixel. For instance, a vertical patch might be 100x20 pixels, while a horizontal patch is 20x100 pixels. This "crossbar" configuration allows the network to incorporate long-range spatial information along the primary axes—helping the model understand the tumor's global structure—while maintaining a narrow width to exclude irrelevant diagonal noise.
Cascaded Training Strategy
The network consists of two sub-models: a horizontal sub-model and a vertical sub-model. These are trained in a cascaded, iterative manner. This strategy is designed to handle "hard" samples—pixels that are difficult to classify correctly. If the vertical sub-model fails to accurately segment a region, that region is prioritized during the training of the horizontal sub-model. This feedback loop ensures that the sub-models complement each other, leading to a more robust final segmentation through a majority voting process at inference time.
Experimental Results and Evaluation
The performance of Crossbar-Net was validated across multiple medical imaging datasets, demonstrating its versatility and accuracy.
Kidney Tumor Segmentation
On a dataset of 3,500 CT slices, Crossbar-Net achieved a Dice Similarity Coefficient (DSC) of 0.913, significantly outperforming standard architectures like U-Net (0.838) and SegCaps. It also recorded a low Hausdorff distance of 8.891 mm, indicating high precision in boundary detection.
Generalization to Other Tasks
To test the model's robustness, it was applied to breast mass segmentation (DDSM dataset) and cardiac MR segmentation. In the breast mass task, it achieved a DSC of 0.9122. In cardiac segmentation, it effectively delineated the left ventricle and myocardium, proving that the crossbar patch concept is applicable beyond kidney imaging.
Computational Efficiency
Despite the cascaded training rounds, Crossbar-Net remains computationally efficient. Training takes approximately one hour, and the inference time for a full patient scan (roughly 35 slices) is less than 1.5 seconds on a standard GPU, making it suitable for clinical workflows.
Conclusion
Crossbar-Net introduces a simple yet powerful modification to the standard CNN training pipeline. By using orthogonal rectangular patches and a collaborative cascaded training strategy, the model effectively balances local detail with global context. The results demonstrate state-of-the-art performance in kidney tumor segmentation and strong generalization to other medical imaging domains, offering a promising tool for automated clinical diagnosis.
Summary updated