PaperlayerView exact revision on arXiv

Source-linked AI summary

DoubleU-Net: A Deep Convolutional Neural Network for Medical Image Segmentation

Debesh Jha, Michael A. Riegler, Dag Johansen, Pål Halvorsen, Håvard D. Johansen

arXiv:2006.04868v2eess.IVcs.CV

TL;DR

Medical image segmentation needs accurate and generalizable pixel-wise prediction despite limited and variable data. The paper proposes DoubleU-Net, a two-U-Net architecture combining pre-trained VGG-19, ASPP, and feature recalibration, and reports improved performance across four datasets. The authors present it as a baseline for medical segmentation and cross-dataset evaluation, while noting higher parameter count and training time than U-Net.

  • Problem

    Medical image segmentation faces limited annotated data, image-quality variation, protocol variation, and patient-image variation, creating a need for automatic and generalizable approaches.

  • Method

    DoubleU-Net stacks two U-Net networks, using pre-trained VGG-19 in the first encoder, ASPP, and squeeze-and-excite blocks.

  • Results

    DoubleU-Net reports significantly better performance than baselines and U-Net across four medical image segmentation datasets.

  • Takeaways & Limitations

    The authors propose DoubleU-Net as a baseline for medical image segmentation and cross-dataset evaluation of generalizability.

  • Takeaways & Limitations

    Compared with U-Net, DoubleU-Net uses more parameters, increasing training time.

Abstract

from arXiv · show

Semantic image segmentation is the process of labeling each pixel of an image with its corresponding class. An encoder-decoder based approach, like U-Net and its variants, is a popular strategy for solving medical image segmentation tasks. To improve the performance of U-Net on various segmentation tasks, we propose a novel architecture called DoubleU-Net, which is a combination of two U-Net architectures stacked on top of each other. The first U-Net uses a pre-trained VGG-19 as the encoder, which has already learned features from ImageNet and can be transferred to another task easily. To capture more semantic information efficiently, we added another U-Net at the bottom. We also adopt Atrous Spatial Pyramid Pooling (ASPP) to capture contextual information within the network. We have evaluated DoubleU-Net using four medical segmentation datasets, covering various imaging modalities such as colonoscopy, dermoscopy, and microscopy. Experiments on the MICCAI 2015 segmentation challenge, the CVC-ClinicDB, the 2018 Data Science Bowl challenge, and the Lesion boundary segmentation datasets demonstrate that the DoubleU-Net outperforms U-Net and the baseline models. Moreover, DoubleU-Net produces more accurate segmentation masks, especially in the case of the CVC-ClinicDB and MICCAI 2015 segmentation challenge datasets, which have challenging images such as smaller and flat polyps. These results show the improvement over the existing U-Net model. The encouraging results, produced on various medical image segmentation datasets, show that DoubleU-Net can be used as a strong baseline for both medical image segmentation and cross-dataset evaluation testing to measure the generalizability of Deep Learning (DL) models.

I. INTRODUCTION

The paper motivates a robust, generalizable medical image segmentation approach and proposes DoubleU-Net, combining two U-Nets with VGG-19 and ASPP. Evaluation across four medical imaging datasets reports improved segmentation over baseline algorithms and U-Net.

  • Medical image segmentation labels each pixel of an object of interest and supports applications including diagnosis and therapy planning.
  • Limited annotated data, image quality, segmentation protocols, and patient-image variation motivate automatic, generalizable, and efficient approaches.
  • DoubleU-Net sequences two U-Nets with two encoders and decoders, using pre-trained VGG-19 in the first encoder and ASPP.
  • The study evaluates DoubleU-Net on four datasets spanning colonoscopy, dermoscopy, and microscopy.
  • DoubleU-Net shows better segmentation performance than baseline algorithms across the 2015 MICCAI, CVC-ClinicDB, ISIC-2018 Lesion Boundary Segmentation, and 2018 Data Science Bowl datasets.
  • An extensive four-dataset evaluation reports significant improvement over U-Net and proposes DoubleU-Net as a new medical image segmentation baseline.

II. RELATED WORK

Prior work established deep CNN architectures for natural and medical image segmentation, while continued challenges in medical imaging include limited test data, imbalance, and difficult cases. DoubleU-Net is proposed to address the need for more accurate segmentation of challenging medical images.

  • Earlier segmentation research developed fully convolutional, dilated, residual, dense, multi-resolution, and nested U-Net architectures.
  • ResUNet++ integrates squeeze-and-excite, ASPP, and attention blocks and reports improved output masks compared with U-Net and ResUNet on two datasets.
  • Recent studies increasingly test models across datasets because generalizability has become a focus of segmentation research.
  • Medical imaging remains challenging because test datasets are scarce, datasets are often imbalanced, and difficult images such as flat polyps may be missed.
  • DoubleU-Net is introduced to produce more accurate segmentation for challenging medical images.

III. THE DOUBLEU-NET ARCHITECTURE

DoubleU-Net stacks two modified U-Nets: the first uses VGG-19, ASPP, and squeeze-and-excite components, while the second refines the first network’s masked input to produce another segmentation mask.

  • DoubleU-Net begins with a VGG-19 encoder followed by a decoder, incorporating ASPP and squeeze-and-excite blocks.
  • The second network differs from U-Net mainly through its use of ASPP and squeeze-and-excite blocks.
  • Network 1 produces Output1, whose element-wise product with the input image becomes the input to Network 2, producing Output2.
  • The two masks are concatenated to compare the intermediate and final predictions qualitatively.
  • The design assumes that reusing the input image with its first mask can improve the subsequent segmentation, while squeeze-and-excite reduces redundant information and ASPP extracts high-resolution feature maps.

A. Encoder Explanation

The architecture uses a pre-trained VGG-19 first encoder and a second encoder built from scratch, with convolution, normalization, activation, feature recalibration, and pooling operations. Two modified decoders upsample features and use encoder skip connections to preserve spatial information.

  • Encoder Explanation: The first encoder uses pre-trained VGG-19, whereas the second encoder is built from scratch.
  • Encoder Explanation: Each second-encoder block applies two 3 × 3 convolutions followed by batch normalization.
  • Encoder Explanation: ReLU introduces nonlinearity, squeeze-and-excite enhances feature-map quality, and 2 × 2 max-pooling with stride 2 reduces spatial dimensions.
  • Encoder Explanation: The network uses two decoders with 2 × 2 bilinear upsampling and concatenated encoder skip connections.
  • Encoder Explanation: The first decoder uses skip connections from the first encoder, while the second uses connections from both encoders.

A. Datasets

The study evaluates DoubleU-Net on four publicly available biomedical segmentation datasets spanning polyp, skin-lesion, and nuclei segmentation.

  • A. Datasets: Four publicly available medical datasets were used to evaluate DoubleU-Net.The datasets include the 2015 MICCAI polyp-detection challenge, CVC-ClinicDB, ISIC-2018 Lesion Boundary Segmentation, and the 2018 Data Science Bowl nuclei dataset.
  • A. Datasets: The 2015 MICCAI challenge uses CVC-ClinicDB for training and ETIS-Larib for testing polyp detection.
  • A. Datasets: CVC-ClinicDB supports comparison for polyp segmentation, while ISIC-2018 provides skin lesions with corresponding annotations.
  • A. Datasets: The 2018 Data Science Bowl dataset provides a publicly available nuclei-segmentation benchmark.
  • A. Datasets: The datasets are clinically relevant because their segmentation can be important for diagnosis and patient outcomes.

B. Evaluation metrics

DoubleU-Net is evaluated with four segmentation metrics across all datasets, while comparisons emphasize each challenge’s official metric.

  • B. Evaluation metrics: DoubleU-Net is evaluated using DSC, mIoU, Precision, and Recall on all four datasets.
  • B. Evaluation metrics: Challenge comparisons emphasize the official evaluation metric, such as mIoU for the Lesion Boundary Segmentation challenge.
  • B. Evaluation metrics: The experiments use an 80% training, 10% validation, and 10% testing split for every dataset.

D. Data augmentation techniques

The supplied passages describe dataset augmentation and evaluation procedures, alongside qualitative polyp results used to assess DoubleU-Net.

  • D. Data augmentation techniques: Medical datasets are split into training, validation, and testing sets before augmentation is applied.
  • D. Data augmentation techniques: Different augmentation methods, including center crop and random transformations, are applied to each dataset split.
  • D. Data augmentation techniques: Figure 2 presents qualitative DoubleU-Net results for large, medium, and flat polyps from the 2015 MICCAI challenge.
  • D. Data augmentation techniques: A single image was converted into 25 augmented images, producing 26 images including the original.
  • D. Data augmentation techniques: The same augmentation techniques were applied to all four datasets.

A. Comparison on 2015 MICCAI sub-challenge on automatic polyp detection dataset

The supplied results passages report DoubleU-Net performance across polyp, skin-lesion, and nuclei datasets, with gains over several baselines and stronger masks on challenging images.

  • A. Comparison on 2015 MICCAI sub-challenge on automatic polyp detection dataset: DoubleU-Net achieved a DSC of 0.7649 and a mIoU of 0.6255 on the 2015 MICCAI polyp-detection dataset.
  • A. Comparison on 2015 MICCAI sub-challenge on automatic polyp detection dataset: DoubleU-Net achieved a DSC of 0.9239 and a mIoU of 0.8611 on CVC-ClinicDB.

C. Comparison on Lesion Boundary segmentation challenge dataset

On the Lesion Boundary Segmentation challenge dataset, DoubleU-Net achieved strong segmentation results and outperformed U-Net and Multi-ResUNet in mIoU.

  • DoubleU-Net achieved a DSC of 0.8962 and mIoU of 0.8212 on the challenge dataset.
  • DoubleU-Net outperformed U-Net by approximately 5.7% in mIoU on the ISIC-2018 lesion boundary dataset.
  • DoubleU-Net outperformed Multi-ResUNet by approximately 1.83% in mIoU on the same dataset.
  • The final segmentation output was better than the intermediate output across small, medium, and large lesions.
  • DoubleU-Net was proposed as a new baseline for semantic image segmentation.

VI. DISCUSSION

The discussion reports that DoubleU-Net performs well across datasets and challenging images, with transfer learning improving cross-dataset performance and compensating for limited training data.

  • DoubleU-Net showed significant DSC improvements over U-Net on the 2015 MICCAI and 2018 Data Science Bowl datasets.The reported improvements were 0.4729% and 15.60%, respectively.
  • DoubleU-Net outperformed competitors in the cross-dataset generalization test on the 2015 MICCAI challenge.
  • A pre-trained ImageNet model performed much better than a model trained from scratch on the cross-dataset test.
  • DoubleU-Net produced better segmentation masks on challenging images and high-quality masks across multi-organ and multi-center medical images.
  • Transfer learning from a pre-trained ImageNet network significantly improved results on every dataset.The authors relate this improvement to compensating for insufficient training data.

VII. CONCLUSION

The conclusion presents DoubleU-Net as a CNN architecture combining two U-Nets with VGG-19, squeeze-and-excite, and ASPP components, while noting opportunities for refinement and a parameter-related training cost.

  • DoubleU-Net combines two U-Net networks, VGG-19, a squeeze-and-excite block, and ASPP.
  • DoubleU-Net performed significantly better than baseline models and U-Net on all four datasets.
  • The architecture can integrate other CNN blocks, and post-processing methods may further improve segmentation results.
  • DoubleU-Net uses more parameters than U-Net, increasing training time.
  • Future work includes simplifying the architecture while retaining its ability to produce high-quality segmentation masks.
Loading 2006.04868v2…