PaperlayerView exact revision on arXiv

Source-linked AI summary

Boundary-weighted Domain Adaptive Neural Network for Prostate MR Image Segmentation

Qikui Zhu, Bo Du, Pingkun Yan

arXiv:1902.08128v2cs.CV

TL;DR

Prostate MR segmentation is challenged by weak boundaries, complex appearance, and limited annotated data. The paper proposes BOWDA-Net, which combines boundary-weighted losses with adversarial domain adaptation, and reports more accurate boundaries and superior results on prostate MR datasets.

  • Problem

    Weak boundaries, complex image appearance, and insufficient annotated data limit accurate prostate MR segmentation and training of complex CNNs.

  • Method

    BOWDA-Net combines boundary-weighted transfer and segmentation losses with adversarial domain adaptation and deep supervision.

  • Results

    The method reports more accurate boundaries and superior results compared with other state-of-the-art methods on PROMISE12 and BWH datasets.

  • Takeaways & Limitations

    Boundary-weighted knowledge transfer and segmentation provide a paper-supported approach for improving prostate MR segmentation under weak-boundary and limited-data conditions.

Abstract

from arXiv · show

Accurate segmentation of the prostate from magnetic resonance (MR) images provides useful information for prostate cancer diagnosis and treatment. However, automated prostate segmentation from 3D MR images still faces several challenges. For instance, a lack of clear edge between the prostate and other anatomical structures makes it challenging to accurately extract the boundaries. The complex background texture and large variation in size, shape and intensity distribution of the prostate itself make segmentation even further complicated. With deep learning, especially convolutional neural networks (CNNs), emerging as commonly used methods for medical image segmentation, the difficulty in obtaining large number of annotated medical images for training CNNs has become much more pronounced that ever before. Since large-scale dataset is one of the critical components for the success of deep learning, lack of sufficient training data makes it difficult to fully train complex CNNs. To tackle the above challenges, in this paper, we propose a boundary-weighted domain adaptive neural network (BOWDA-Net). To make the network more sensitive to the boundaries during segmentation, a boundary-weighted segmentation loss (BWL) is proposed. Furthermore, an advanced boundary-weighted transfer leaning approach is introduced to address the problem of small medical imaging datasets. We evaluate our proposed model on the publicly available MICCAI 2012 Prostate MR Image Segmentation (PROMISE12) challenge dataset. Our experimental results demonstrate that the proposed model is more sensitive to boundary information and outperformed other state-of-the-art methods.

I. INTRODUCTION

Prostate MR segmentation is clinically valuable but remains difficult because boundaries are weak, images are complex, and annotated training data are limited. BOWDA-Net addresses these challenges through boundary-weighted losses and transfer learning.

  • Prostate MR segmentation supports prostate disease diagnosis and treatment, particularly for prostate cancer.
  • Manual segmentation is expensive, time-consuming, and prone to inter- and intra-observer variation, motivating automated methods.
  • Weak prostate boundaries, complex background texture, and variation in size, shape, and intensity complicate accurate segmentation.
  • Limited annotated training data make it difficult to fully train complex CNNs for prostate segmentation.
  • BOWDA-Net uses boundary-weighted transfer and segmentation losses with deep supervision to improve boundary-sensitive segmentation and exploit additional datasets.

II. RELATED WORKS

Related work spans shape- and appearance-based prostate segmentation, CNN methods, and domain adaptation. The review emphasizes persistent domain-shift and weak-boundary challenges when models encounter new medical-image datasets.

  • Conventional prostate segmentation methods use shape priors, image priors, atlases, deformable models, graph cuts, and active contours.
  • Existing approaches still face weak or ambiguous prostate boundaries and variability in image contrast and appearance.
  • CNN-based approaches improve feature representation but may combine learned features with traditional segmentation models such as active shape models or level sets.
  • Domain shift can degrade CNN segmentation performance on datasets that differ from the training data.
  • Medical-image domain adaptation methods address scanner, modality, sampling, and cross-patient differences using adversarial or reverse-adaptation strategies.

III. MATERIALS

The experiments use PROMISE12 as the target domain, a Philips 3T dataset as the source domain, and an additional BWH multi-parametric MR dataset.

  • PROMISE12 contains 50 transversal T2-weighted prostate MR images with corresponding ground-truth segmentations acquired in different hospitals.
  • The source dataset contains 81 prostate MR volumes acquired with a Philips 3T scanner and an endorectal coil.
  • Each Philips source volume has approximately 26 slices of 512×512 pixels, with 0.27mm×0.27mm in-plane resolution and 3mm inter-plane distance.
  • t-SNE visualization of 280 slices from each source and target domain reveals domain shift, while boundary gradient distributions are also compared.
  • The BWH dataset comprises baseline and repeat multi-parametric prostate MR exams for 15 subjects acquired within two weeks.

IV. BOUNDARY-WEIGHTED DOMAIN ADAPTATION

BOWDA-Net adapts prostate segmentation across source and target domains with dual segmentation networks and a domain discriminator. Boundary-weighted transfer and segmentation losses focus adaptation and supervision on difficult boundaries.

  • BOWDA-Net contains source and target segmentation networks, SNet-s and SNet-t, plus a domain feature discriminator D.
  • SNet-s and SNet-t learn source- and target-domain representations, while D distinguishes their extracted features in an adversarial training process.
  • The source network is trained with supervised cross entropy, frozen, and used to initialize the target network before end-to-end adaptation.
  • The discriminator output gives each spatial feature-map unit the probability that its corresponding input patch belongs to the target domain.
  • Boundary-weighted transfer loss makes domain adaptation emphasize boundaries, while boundary-weighted segmentation loss supervises segmentation-network training toward accurate boundaries.
  • Boundary maps are generated from ground-truth labels with a Sobel filter and smoothed using a 3×3 Gaussian filter with variance σ2 = 0.64.

B. Boundary-weighted Segmentation Loss

The paper introduces a boundary-weighted segmentation loss to improve sensitivity to prostate boundaries, which cross entropy alone may inadequately represent. Its distance-based term regularizes segmentation position, shape, and continuity toward ground-truth boundaries.

  • Cross entropy can compromise boundary accuracy because whole-region information may overwhelm boundary information.
  • BWSL adds a distance loss to make the segmentation network more sensitive to boundaries.
  • The distance loss regularizes segmentation position, shape, and continuity so predictions remain close to object boundaries.
  • The distance map is constructed from a distance transform of ground-truth boundaries, with β controlling the weighting coefficient.

C. Network Design and Configurations

SNet is a 3D volumetric segmentation network inspired by U-Net and DenseNet, combining down- and up-sampling paths with densely connected residual blocks. Its DRBs target small-data overfitting and information propagation while retaining compactness and hierarchical representation.

  • SNet is a 3D network designed to leverage volumetric spatial context for prostate image segmentation, drawing inspiration from U-Net and DenseNet.
  • SNet contains down-sampling and up-sampling paths, with convolutional blocks, densely-connected residual blocks, pooling layers, and deconvolutional layers.
  • DRBs combine dense connections, transition layers, and residual connections to address overfitting with small datasets and promote faster information propagation.
  • Dense feature concatenation, transition-layer fusion, and residual connections help DRBs become deeper with fewer parameters and hierarchical representation capability.
  • SNet exceeds 100 layers and uses specified DRB configurations, growth rate 32, and dropout rate 0.3 to mitigate overfitting.

V. EXPERIMENTS

The experiments standardize PROMISE12 volumes through resampling and intensity normalization, while normalizing Philips 3T volumes without resampling. Testing uses overlapping sliding windows whose probability outputs are averaged into the final segmentation.

  • PROMISE12 volumes are resampled to 0.625mm×0.625mm×1.5mm and normalized to zero mean and unit variance.
  • Philips 3T volumes with 0.27mm×0.27mm×3mm resolution are intensity-normalized to zero mean and unit variance.
  • Testing crops target images into overlapping 16×96×96 sub-volumes with stride 8×48×48.
  • Overlapping sub-volume probability maps are averaged to produce the final volume segmentation.
  • Models are trained end-to-end with SGD using learning rate 0.0001, momentum 0.9, weight decay 1.0×10−6, and batch size 4.

B. Segmentation Performance

BOWDA-Net ranked first on the PROMISE12 challenge evaluation and produced accurate prostate contours in qualitative examples.

  • PROMISE12 evaluation used 30 testing MR images with held-out ground truth, and boundary measures were calculated for the apex and base.The apex and base were defined as the first and last approximately one-third of the prostate along the axial direction.
  • Table I reports quantitative evaluation results for BOWDA-Net and other methods on the PROMISE12 challenge dataset.
  • 89.59 overall score ranked BOWDA-Net first among the listed PROMISE12 challenge teams.The reported results were obtained from the challenge website on Jan 21, 2019.
  • Qualitative examples showed accurate segmentations that delineated clear prostate contours in MR images.Figure 4 compares yellow ground-truth contours with red contours from the proposed segmentation.

C. Impact of Loss Function

The loss-function experiments found that boundary-weighted transfer and segmentation losses improved segmentation performance and produced smoother, more accurate prostate boundaries.

  • The ablation study compared supervised and adversarial losses, including Lce, LSeg, LD, BWTL, and BWSL.
  • The training-strategy experiments used PROMISE12 target data and Philips 3T MR source data, with target data split into 10 validation subjects and 40 training subjects.
  • Using LSeg and LD improved performance over Lce, while combining BWTL and BWSL achieved the best results on most evaluation metrics.The authors interpret this as improved effectiveness in securing prostate boundaries.
  • BWTL and BWSL produced smoother and more accurate boundaries than the alternative loss settings.Figure 5 presents examples comparing segmentation results obtained with different loss functions.
  • 89.81% DSC after resampling source data exceeded the 87.78% obtained without resampling, but direct mixing still suffered from domain shift and source-data imbalance.
  • BOWDA-Net significantly outperformed the other evaluated training strategies with p < 0.05 in paired statistical comparisons.The comparison used a paired t-test with a confidence interval of 0.95.

E. Comparison with Other Transfer Learning Methods

The study compares BOWDA-Net with alternative training strategies, VNet-based approaches, and a published domain adaptation method using matched prostate MR datasets. The comparison is organized around PROMISE12 results and quantitative evaluations across the listed methods.

  • Comparison setup: The comparison uses the same 81+50 cases from the Philips 3T MR and PROMISE12 datasets for fair evaluation.The selected published method was highly ranked and had publicly available source code.
  • Compared methods: BOWDA-Net is compared with a domain adaptation method combining a segmentation FCN and a CNN domain classifier.The comparison uses the same training and test datasets and target-domain labels for model training.
  • Training strategies: VNet is evaluated with direct source-target mixing and mixing after resampling the source data to the target resolution.BOWDA-Net instead uses source-domain images without resampling them to match the target data.
  • Results: The segmentation results for BOWDA-Net and the compared methods are reported in Table IV for the PROMISE12 challenge dataset.Figure 6 presents box plots of segmentation evaluation results across different training strategies and networks.

F. Evaluation on the BWH Dataset

The BWH evaluation uses Philips 3T MR images as source data and evaluates the proposed model under four training settings. BOWDA-Net achieves the highest reported DSC and significantly outperforms the other methods.

  • Dataset and protocol: The BWH experiment uses 81 Philips 3T prostate MR images as source data and three randomly selected BWH subjects as test data.The remaining BWH subjects are used for training.
  • Results: BOWDA-Net achieves the highest DSC of 89.67% on the BWH dataset.Table V reports segmentation performance for the training strategies described in the experiment.
  • Statistical comparison: BOWDA-Net significantly outperforms the other methods on the BWH dataset with p < 0.05.The statistical comparison is reported alongside the segmentation results in Table V.

G. Network Ablation Study

The ablation study evaluates the contribution of dense, residual, and long connections within the proposed segmentation network. Adding these connections produces more accurate segmentation than the alternative network configurations.

  • Ablation configurations: Four configurations are compared: FCN, FCN + Dense, FCN + Dense + Residual, and the proposed SNet with dense, residual, and long connections.The configurations are trained on PROMISE12 target-domain data only.
  • Network design: The proposed SNet combines dense, residual, and long connections in the domain image segmentation network.The other configurations progressively omit one or more of these connection types.
  • Ablation result: Adding residual, long, and dense connections helps achieve more accurate segmentation than the other networks.This finding is reported for the PROMISE12 ablation comparison.
Loading 1902.08128v2…