Crossbar-Net: A Novel Convolutional Network for Kidney Tumor Segmentation in CT Images

I. INTRODUCTION

• Accurate segmentation supports tumor quantification and classification, while manual delineation is undesirable because it can be subjective and must cover many images.The paper describes automatic CT kidney-tumor segmentation as being in high demand.Source: §I · p. 1
• Existing methods can fail when tumors resemble surrounding tissue or have internally dissimilar intensity, leaving accurate segmentation an ongoing need.The introduction contrasts easier, high-contrast cases with tumors connected to or visually similar to surrounding tissue.Source: §I · p. 2
• Crossbar patches span tumors horizontally or vertically, combining local detail with contextual and symmetrical information that squared patches may miss.The authors note that simply enlarging square patches or using whole images may introduce irrelevant noise and obscure boundary details.Source: §I · p. 2
• Crossbar-Net trains vertical and horizontal sub-models in cascaded rounds, using each model’s difficult regions to provide complementary training to the other.The paper also describes basic sampling to select discriminative patches and balance classes, and covering re-sampling to support complementary help.Source: §I · p. 2
• The authors report promising generalization on cardiac and breast-mass segmentation, and kidney-tumor training and testing times of about 1 hour and 1.5 seconds on a regular GPU.Crossbar patches require no additional parameters to train.Source: §I · p. 3

II. RELATED WORK

• Earlier kidney-tumor methods using energy minimization or hand-crafted features perform well in simpler cases, but are less assured when tumor appearance resembles surrounding tissue.Examples include region growing, active contours, texture features, and classifiers.Source: §II · p. 3
• Medical-image CNNs borrowed for kidney-tumor segmentation used squared patches or whole-image inputs, while the authors identify square-patch locality as a limitation for this task.The related work describes both patch-based and image-based alternatives.Source: §II · p. 3
• Unlike prior image-level or squared-patch methods, Crossbar-Net uses non-squared patches to capture local detail and vertical and horizontal context.This distinction is presented as the approach’s feature-representation difference.Source: §II · p. 3
• Unlike other cascaded approaches, Crossbar-Net combines vertical and horizontal sub-models that iteratively improve by complementing one another.The authors describe enhancing regions around misclassified pixels during later rounds.Source: §II · p. 4

A. Our Methodology

• The method converts segmentation into pixel-wise classification, predicting whether a patch belongs to the tumor or non-tumor class.The framework includes both training and testing stages.Source: A. Our Methodology · p. 4
• At each training round, the method identifies mis-segmented regions and re-samples them for the other directional sub-model, alongside patches from basic sampling.Training repeats until error converges or the maximum number of rounds is reached.Source: A. Our Methodology · p. 4
• For a new CT image, the trained sub-models from each round produce segmentations that are combined by majority vote.Source: A. Our Methodology · p. 4

1) Basic Sampling Strategy:

• Basic sampling uses denser non-tumor patches near the tumor and sparser patches farther away, while sampling tumor patches uniformly.The strategy emphasizes the boundary region considered hard to segment and uses different sampling intervals across rounds.Source: 1) Basic Sampling Strategy · p. 5

2) Covering Re-sampling Strategy:

• Covering re-sampling spans a mis-segmented vertical region with horizontal patches centered on its left, middle, and right lines, producing roughly 40 patches per vertical patch.The process also samples every three pixels along each line to limit redundancy.Source: 2) Covering Re-sampling Strategy · p. 5

C. Sub-Model Architecture

• Both sub-models use eight convolutional layers, two max-pooling layers, and a softmax layer, with the vertical model taking a 100 × 20 patch as input.The vertical model uses non-square convolutional kernels; architecture settings were experimentally determined by inner cross-validation.Source: C. Sub-Model Architecture · p. 5
• The horizontal sub-model takes a 20 × 100 patch, and the architecture may differ across segmentation scenarios.The paper uses the same architecture for cardiac segmentation and a different one for breast mass segmentation.Source: C. Sub-Model Architecture · p. 5

D. Training Sub-Models

• Initial vertical and horizontal sub-models are trained on crossbar patches sampled with the basic strategy, including tumor patches and non-tumor patches centered on contours.Non-tumor sampling intervals are smaller near the tumor boundary than farther away.Source: D. Training Sub-Models · p. 5
• In later rounds, covering re-sampling targets mis-segmented regions, while basic sampling controls the amount and distribution of samples and avoids duplicate patches.The two sampling strategies are combined to generate new patches for the sub-models.Source: D. Training Sub-Models · p. 6
• Training repeats updates to the horizontal and vertical sub-models until the maximum round count is reached or their training errors no longer decrease significantly.Source: D. Training Sub-Models · p. 6
• Vertical and horizontal sub-models complement one another during cascaded training, with patches from mis-segmented regions used to emphasize those regions in subsequent rounds.The authors note that boundaries blurred in one direction may be sharp in the other.Source: D. Training Sub-Models · p. 6
• The authors describe self-improvement across rounds as potentially arising from newly sampled training data and increased emphasis on regions both sub-models misclassify.They characterize the expected improvement on those regions as analogous to AdaBoost’s weight-updating approach.Source: D. Training Sub-Models · p. 6

• For each pixel in a new image, crossbar patches are extracted and passed to the trained vertical and horizontal sub-models in each round.Source: E. Testing · p. 6
• Each sub-model predicts whether the central pixel of its patch belongs to the tumor region.Source: E. Testing · p. 6
• The final segmentation is determined by majority voting across the 2T sub-models produced over T rounds.Source: E. Testing · p. 6

IV. EXPERIMENTAL RESULTS

• The experiments examine Crossbar-Net’s characteristics, compare it with baseline methods on kidney tumors, and apply it to cardiac and breast mass segmentation.Source: §IV · pp. 6–7

A. Data

• The kidney tumor dataset contains 3,500 CT slices from 94 subjects, with one tumor per slice and physician annotations.The data were divided by subject into training, validation, and testing sets of 50, 8, and 36 subjects.Source: A. Data · p. 7
• The breast mass experiments use 1,000 selected DDSM images, split into 600 training, 100 validation, and 300 testing images.Experienced radiologists annotated tumor boundaries, and image regions outside the image in extracted patches are filled with black.Source: A. Data · p. 7
• The cardiac dataset comprises 7,980 MRI slices from 33 subjects, with 20 subjects for training, 3 for validation, and 10 for testing.Ground-truth contours are provided for the left ventricle’s endocardium and epicardium.Source: A. Data · p. 7

B. Evaluation Criteria

• DSC measures overlap, TPF measures correctly predicted tumor-pixel coverage, and CD and HD measure distances between predicted and manual segmentations.Higher DSC and TPF indicate better overlap and coverage; smaller CD and HD indicate closer segmentations.Source: B. Evaluation Criteria · p. 7

C. Implementation Details

• Experiments were repeated three times with randomly selected training, validation, and test subsets, and reported performance was averaged.Sub-models converged within 20 epochs on kidney tumors and 25 epochs on cardiac and DDSM data.Source: C. Implementation Details · p. 7

D. Characteristics of Crossbar-Net

• Separately trained sub-models improved as rounds increased but required 10 rounds to converge, compared with 3 rounds under the proposed joint training manner.The separate-training experiment excluded the other directional sub-model throughout.Source: D. Characteristics of Crossbar-Net · p. 8
• Vertical and horizontal sub-models corrected different tumor regions, and later sub-models generally performed better than earlier ones.Some boundary errors were shared across directions and persisted in early rounds.Source: D. Characteristics of Crossbar-Net · p. 8
• Voting across all sub-models outperformed combining only V3 and H3 by around 2% in DSC and TPF, with V3 and H3 weighted 1.5 and the others 1.As rounds increased, majority voting remained superior to combining only the last two sub-models.Source: D. Characteristics of Crossbar-Net · p. 9
• Squared patches produced much lower DSC and TPF than crossbar patches, and enlarging squared patches to 100 × 100 brought no significant improvement.The authors suggest larger squared patches may add irrelevant noise when distinguishing boundaries.Source: D. Characteristics of Crossbar-Net · p. 9

E. Comparison to Other Methods on Kidney Tumor

• The comparison included hand-crafted methods, multi-scale 2D- and 3D-CNNs, FCN-G, U-Net, V-Net, and SegCaps.The hand-crafted baselines extracted the kidney area, calculated features manually, and then classified or segmented tumors.Source: E. Comparison to Other Methods on… · p. 10
• Testing used 600 randomly sampled images from one test set comprising 30 subjects.The authors used density estimation to depict DSC distributions across these tumors.Source: E. Comparison to Other Methods on… · p. 10
• On 600 sampled kidney tumors, Crossbar-Net achieved DSC above 0.9 for most tumors; the 2D-CNN and hand-crafted methods had many low-DSC cases.U-Net performed better than those methods and the multi-scale 3D-CNN, while SegCaps achieved the second-best performance; U-Net and SegCaps also had some zero-DSC cases.Source: E. Comparison to Other Methods on… · p. 10
• Crossbar-Net achieved higher DSC and TPF and smaller HD and CD than the other compared methods on kidney tumors.Table III reports average values of all four metrics across test sets.Source: E. Comparison to Other Methods on… · p. 10, Table III · p. 10
• Crossbar-Net, including six sub-models, trained in about 1 hour on one GPU and segmented a new patient in less than 1.5 seconds.The reported patient had about 35 slices; U-Net was very close to Crossbar-Net in training and testing time.Source: E. Comparison to Other Methods on… · p. 12

F. Crossbar-Net for Breast Mass Segmentation

• DDSM sub-models again showed self-improvement and mutual help, consistent with the characteristics observed in kidney experiments.The passage identifies the breast-mass results as confirming these characteristics.Source: F. Crossbar-Net for Breast Mass… · p. 12
• Crossbar-Net was slightly superior to the other methods in breast-mass segmentation on DDSM.The comparison reports DSC values for Crossbar-Net and several other methods.Source: F. Crossbar-Net for Breast Mass… · p. 12
• All six sub-models took about 6 hours to train and about 15 seconds to segment a new DDSM subject.The authors attribute the higher cost than on kidney and cardiac datasets to larger patches and a larger segmentation model.Source: F. Crossbar-Net for Breast Mass… · p. 12

G. Crossbar-Net for Cardiac Segmentation

• LV sub-model performance improved gradually across rounds in the same direction.The reported experiment segmented LV and myocardium in cardiac MRI.Source: G. Crossbar-Net for Cardiac Segmentation · p. 12
• For LV segmentation, Crossbar-Net had the second-best DSC overall, while its HD and AD were superior to those of the other compared methods.GridNet had the most accurate LV DSC; for MYO segmentation, Crossbar-Net outperformed all other compared methods.Source: G. Crossbar-Net for Cardiac Segmentation · p. 13
• The authors report better performance on cardiac segmentation than on kidney tumors because the LV cavity and myocardium have more regular shapes.Source: G. Crossbar-Net for Cardiac Segmentation · p. 13

V. CONCLUSION

• Crossbar-Net combines crossbar patches, basic and covering re-sampling, and cascaded training as its three main design innovations.The method is reported to handle kidney tumors with varied size, shape, contrast, and appearance.Source: §V · p. 13
• Basic and covering re-sampling enhance mis-segmented regions while preventing either sub-model from becoming over-emphasized on them.The directional sub-models can complement each other’s segmentation and improve themselves through re-sampling.Source: §V · p. 13
• Reported applications to cardiac and breast mass segmentation support the paper’s claim that Crossbar-Net has a wide range of application.The authors identify extending symmetric information beyond horizontal and vertical axes as future work.Source: §V · pp. 13–14
• Future work would extend symmetric information from horizontal and vertical axes to other axes.Source: §V · p. 14