PaperlayerView exact revision on arXiv

Source-linked AI summary

Rethinking the Unpretentious U-net for Medical Ultrasound Image Segmentation

Gongping Chen, Lei Li, JianXun Zhang, Yu Dai

arXiv:2209.07193v4eess.IVcs.CVcs.LG

TL;DR

Breast tumor segmentation is difficult because ultrasound tumors have variable morphology, blurred boundaries, and similar surrounding intensities, while existing U-net variants can underuse benchmark characterization and add complexity. NU-net combines U-nets of different depths with shared weights and nested feature processing. The authors report competitive breast-tumor results and best performance on the evaluated renal-ultrasound task, without significantly increasing computational cost.

  • Problem

    Variable tumor morphology, blurred boundaries, and similar intensity distributions challenge breast tumor segmentation, while existing variants may overlook benchmark characterization and add reproducibility complexity.

  • Method

    NU-net combines a fifteen-layer backbone U-net, multi-out U-net modules, and multi-step down-sampling short-connections using nested U-nets with shared weights.

  • Results

    NU-net achieves competitive segmentation results on BUSI and Dataset B and the best performance among compared methods for renal ultrasound segmentation.

  • Takeaways & Limitations

    The proposed design improves robustness across breast-tumor scales and transfers favorably to renal ultrasound segmentation while remaining easy to reproduce.

Abstract

from arXiv · show

Breast tumor segmentation is one of the key steps that helps us characterize and localize tumor regions. However, variable tumor morphology, blurred boundary, and similar intensity distributions bring challenges for accurate segmentation of breast tumors. Recently, many U-net variants have been proposed and widely used for breast tumors segmentation. However, these architectures suffer from two limitations: (1) Ignoring the characterize ability of the benchmark networks, and (2) Introducing extra complex operations increases the difficulty of understanding and reproducing the network. To alleviate these challenges, this paper proposes a simple yet powerful nested U-net (NU-net) for accurate segmentation of breast tumors. The key idea is to utilize U-Nets with different depths and shared weights to achieve robust characterization of breast tumors. NU-net mainly has the following advantages: (1) Improving network adaptability and robustness to breast tumors with different scales, (2) This method is easy to reproduce and execute, and (3) The extra operations increase network parameters without significantly increasing computational cost. Extensive experimental results with twelve state-of-the-art segmentation methods on three public breast ultrasound datasets demonstrate that NU-net has more competitive segmentation performance on breast tumors. Furthermore, the robustness of NU-net is further illustrated on the segmentation of renal ultrasound images. The source code is publicly available on https://github.com/CGPzy/NU-net.

1 INTRODUCTION

Breast ultrasound tumor segmentation is clinically important but difficult because tumors vary in morphology, boundaries are blurred, and surrounding intensities can be similar. NU-net addresses limitations of existing U-net variants by combining deeper and nested U-nets with simpler operations.

  • Breast lesion segmentation supports tumor diagnosis and postoperative follow-up, but complex ultrasound patterns and ambiguous boundaries challenge automation.
  • Existing U-net variants add multi-scale, attention, deep-supervision, or hybrid strategies, yet may underuse the benchmark network’s characterization ability.
  • Additional operations also make networks harder to understand and reproduce, motivating a simpler architecture.
  • NU-net uses a fifteen-layer U-net baseline, embeds a multi-out U-net, and adds three short-connections based on multi-step down-sampling.
  • NU-net is designed to improve segmentation robustness and adaptability across tumor scales without significantly increasing computational cost.

2. RELATED WORK

Related work develops U-net variants through multi-scale processing, attention, deep supervision, and hybrid combinations. These strategies can improve segmentation, but their adaptability, computational cost, or complexity remains constrained in some cases.

  • 2.1 Multi-Scale U-net: Multi-scale U-nets use different receptive fields, while SKU-net and AAU-net adaptively select tumor features across scales.
  • 2.1 Multi-Scale U-net: AAU-net selects features across channel and spatial dimensions, but this operation is computationally expensive and variable tumor morphology remains challenging.
  • 2.2 Attention-Optimized U-net: Attention-based U-nets integrate attention modules, although deeper dilated convolutions may fail to capture sufficient contextual information.
  • 2.3 Deep-Supervised U-net: Deep-supervised U-nets constrain intermediate feature maps toward the ground truth, but excessive supervision can increase parameters without improving performance.
  • 2.4 Multi-Module Hybrid U-net: Hybrid designs jointly combine residual learning, multi-scale inputs, attention, and deep supervision, while too many low-level inputs can weaken high-level semantic characterization.

3 METHOD

NU-net is a nested architecture built from U-nets of different depths with shared weights. It combines a deeper backbone, multi-out feature refinement, and multi-step down-sampling connections to improve feature characterization and long-range information exchange.

  • 3 METHOD: NU-net can be understood as seven U-nets with different depths and shared weights, supporting a simple and reproducible design.
  • 3.2 Deeper Backbone U-net: The method uses a fifteen-layer U-net as its baseline because shallower variants may inadequately capture spatial and location information in complex ultrasound images.
  • 3.3 Multi-Out U-net (MOU): The multi-out U-net refines encoded feature maps at depths 11, 9, 7, 5, 3, and 1, then supplies intermediate outputs to the backbone decoder.
  • 3.3 Multi-Out U-net (MOU): MOU outputs provide additional feature information for reconstruction and strengthen correspondence between encoding and decoding features at the same scale.
  • 3.4 Multi-Step Down-Sampling Short-Connection: Multi-step down-sampling is used for short-connections to improve correlation between distant feature maps beyond same-scale or adjacent-scale connections.

4 MATERIALS AND EXPERIMENTS

The study evaluates segmentation on three public breast ultrasound datasets using established metrics, cross-validation, ablation studies, comparisons, robustness analyses, and external validation. Training uses binary cross-entropy with Adam under a fixed implementation setup.

  • 4 MATERIALS AND EXPERIMENTS: Ablation experiments evaluate the deeper U-net, multi-out U-net, and multi-step down-sampling short-connection on BUSI and Dataset B.
  • 4.1 Dataset Description: BUSI contains 780 images from 600 patients, Dataset B contains 163 images, and STU contains 42 images used only for external testing.
  • 4.2 Evaluation Metrics: The evaluation metrics are Jaccard, Precision, Recall, Specificity, and Dice, with higher values indicating better segmentation.
  • 4 MATERIALS AND EXPERIMENTS: Experiments include ablation, state-of-the-art comparison, robustness analysis, and external validation using fold-based cross-validation.
  • 4 MATERIALS AND EXPERIMENTS: Training uses binary cross-entropy and Adam, with an initial learning rate of 0.001, 50 epochs, and batch size 12.

5 RESULTS

NU-net was evaluated through component ablations, comparisons with twelve segmentation methods, and robustness tests on breast ultrasound datasets. Increasing U-net depth and adding nested components improved segmentation while keeping computational-cost growth limited.

  • Architecture ablation: Ablation experiments used four-fold cross-validation on BUSI and Dataset B to assess deeper U-net, MOU, and MDSC components.The baseline U-net contained four down-sampling and four up-sampling operations.
  • Architecture ablation: Depth 15 or 17 produced the best segmentation results, so the 15-layer U-net was selected as the lower-parameter baseline.Increasing depth significantly improved breast-tumor segmentation compared with the original U-net.
  • Architecture ablation: Embedding multi-out U-net modules further improved segmentation accuracy after increasing the baseline network depth.The comparison between Deeper U-net and Deeper U-net + MOU supported feature-map refinement as an effective component.
  • Architecture ablation: Adding network depth and MOU increased parameters without significantly increasing computational cost, while MDSC had little effect on either quantity.Overall, the method achieved more robust segmentation at small computational cost.
  • Comparative evaluation: NU-net achieved the most competitive segmentation results on BUSI and Dataset B among twelve state-of-the-art comparison methods.Paired Student’s t-tests reported significant differences from the comparison methods at p < 0.05.
  • Robustness analysis: NU-net achieved the best performance on benign and malignant tumors and the smallest failure rates, measuring 1.43% and 4.12%, respectively.The benign and malignant evaluations used four-fold and three-fold cross-validation, respectively.

6 APPLICATIONS

NU-net was also evaluated on external breast-ultrasound data and renal ultrasound images. These experiments reported strong segmentation performance and qualitative mitigation of ultrasound-specific disturbances.

  • Renal ultrasound application: NU-net achieved the best segmentation performance among the compared methods on renal ultrasound images.The renal evaluation used the same experimental setup as recent renal-segmentation work and three-fold cross-validation.
  • Renal ultrasound application: The renal evaluation further verified the method’s effectiveness beyond breast-tumor segmentation.The application extended the evaluation to a different ultrasound segmentation task.
  • External validation: External validation used Dataset B and STU to evaluate adaptability across ultrasound data collected from different sites.The experiments addressed differences between sites that can affect performance on external datasets.
  • Renal ultrasound application: The renal-ultrasound results indicate that NU-net can alleviate disturbances from uneven energy distribution, blurred boundaries, and similar surrounding tissue.The qualitative comparison included competing methods, NU-net, and ground-truth contours.

7 DISCUSSIONS

NU-net combines nested U-nets with different depths and shared weights while adding parameters without significantly increasing computational cost. Discussion experiments associate this design with competitive accuracy and robustness across tumor classes, external data, and renal ultrasound.

  • Method discussion: NU-net refines encoded features using U-nets with different depths and reports increased parameters without significantly increasing computational cost.The added MOU and MDSC operations are described as easy to understand and perform.
  • Comparison with U-net variants: On BUSI and Dataset B, NU-net reduced missed and false detections and produced results closer to ground-truth masks across five quantitative metrics.The evaluation used four-fold cross-validation.
  • Robustness analysis: NU-net achieved the best segmentation accuracy for both benign and malignant tumors, indicating adaptability across tumor classes.The discussion also reports minimal failure rates for both classes.
  • Robustness analysis: External experiments found that competing methods could be sensitive to different breast-ultrasound images, whereas NU-net produced masks closer to ground truth.The external evaluation used Dataset B and STU.
  • Statistical test: P-values across Tables 4–7 were almost all below 0.05, supporting reported significant improvements over competing methods on the evaluated metrics.The paper concludes that NU-net has good robustness and generalization ability for breast-tumor segmentation.

8 CONCLUSIONS

The paper identifies recurring limitations in U-net variants for breast ultrasound segmentation and proposes NU-net to address them. NU-net is reported to improve segmentation performance and robustness across breast and renal ultrasound images while remaining simple to reproduce.

  • Existing U-net variants tend to use shallow architectures, introduce complex extra operations, and are inconvenient to reproduce or apply.
  • NU-net combines U-nets with different depths and shared weights, providing a simple architecture that is easy to understand and reproduce.
  • The design reduces sensitivity to input images with different scales and improves the characterization of object or region features.
  • Extensive comparative experiments, robustness analysis, and external validation demonstrate better breast-lesion segmentation performance for NU-net than several state-of-the-art methods.
  • Application to renal ultrasound images further demonstrates the reported robustness of NU-net.
Loading 2209.07193v4…