Source-linked AI summary
Adversarial Deep Structured Nets for Mass Segmentation from Mammograms
Wentao Zhu, Xiang Xiang, Trac D. Tran, Gregory D. Hager, Xiaohui Xie
TL;DR
Mass segmentation supplies morphological features important for diagnosis, while two-stage training can overfit potential functions. The paper proposes an end-to-end adversarial FCN-CRF network with a position prior, and reports improved performance on INbreast and DDSM-BCRP.
Problem
Mass segmentation provides diagnostic morphological features, but two-stage training can produce potential functions that overfit the training data.
Source: §1 · p. 1Method
The paper trains an FCN-CRF end to end with adversarial training and a position prior for mammographic mass segmentation.
Source: §1 · p. 1Results
Takeaways & Limitations
The authors report superior performance for adversarial FCN-CRF on two commonly used public mammographic mass segmentation datasets.
Source: §5 · p. 4
Abstract
from arXiv · showhide
Mass segmentation provides effective morphological features which are important for mass diagnosis. In this work, we propose a novel end-to-end network for mammographic mass segmentation which employs a fully convolutional network (FCN) to model a potential function, followed by a CRF to perform structured learning. Because the mass distribution varies greatly with pixel position, the FCN is combined with a position priori. Further, we employ adversarial training to eliminate over-fitting due to the small sizes of mammogram datasets. Multi-scale FCN is employed to improve the segmentation performance. Experimental results on two public datasets, INbreast and DDSM-BCRP, demonstrate that our end-to-end network achieves better performance than state-of-the-art approaches. \footnote{https://github.com/wentaozhu/adversarial-deep-structural-networks.git}
1. INTRODUCTION
Mass segmentation provides morphological features important for diagnosis, but two-stage training can overfit potential functions. The paper proposes end-to-end adversarial FCN-CRF segmentation with a spatial prior and reports best results on two commonly used mammographic mass datasets.
1. INTRODUCTION
- Mass segmentation provides morphological features that play a crucial role in diagnosis.Source: §1 · p. 1
- Two-stage training can produce potential functions that overfit the training data, while deep-network approaches to mammogram segmentation remain few.Source: §1 · p. 1
- The end-to-end adversarial FCN-CRF framework integrates a spatial prior and achieves the best results on two commonly used mammographic mass segmentation datasets.The network is designed to learn robustly from scarce mammographic images.Source: §1 · p. 1
2. FCN-CRF NETWORK
The network uses an FCN to provide unary potentials and a CRF for structured pixel-label learning. Mean-field inference updates label estimates through message passing, reweighting, compatibility transformation, unary-potential addition, and normalization.
2. FCN-CRF NETWORK
- The FCN is an image-segmentation model, and its training objective is a maximum-likelihood loss over pixel labels.The described images are fixed at 40 × 40 pixels, with 1,600 pixels per image.Source: §2 · pp. 1–2
- The fully connected CRF treats pixel labels as background or mass and combines FCN unary losses with pairwise costs based on label compatibility, learned weights, and Gaussian feature kernels.Source: §2 · p. 2
- Mean-field inference updates pixel-label estimates by message passing, reweighting, compatibility transformation, unary-potential addition, and normalization.The update can be interpreted as a recurrent neural network.Source: §2 · p. 2
- Figure 2 presents empirical prior estimates for INbreast and DDSM-BCRP and shows ground truth alongside width-2 and width-3 trimaps.Source: Figure 2 · p. 2
3. ADVERSARIAL FCN-CRF NETS
The method adds a position-dependent prior to FCN predictions and uses adversarial training with a CRF to model structured pixel labels. Multi-scale FCNs provide the CRF’s unary potentials, and training combines adversarial and empirical losses.
3. ADVERSARIAL FCN-CRF NETS
- The position prior addresses spatially varying mass labels by adding an image-sized bias to FCN softmax predictions, with the resulting negative log probability serving as the CRF unary potential.Masses are observed more often near the ROI center, while its boundary areas are more likely to be background.Source: §3 · p. 2
- Multi-scale FCNs form the unary potential as a learned weighted sum of the potentials supplied by each scale.The learned weights combine the per-scale FCN potentials for use in the CRF.Source: §3 · p. 2
- Adversarial training fits examples under worst-case perturbations that increase loss, using a linear approximation and L2 constraint to calculate the perturbation.The model’s prediction is pixel-independent for adversarial FCN and uses mean-field inference for adversarial CRF.Source: §3 · pp. 2–3
- Training minimizes the sum of adversarial and empirical losses on training samples, with an L2 regularization factor for the network parameters.The probability used in the loss is either the FCN mass prediction or the CRF posterior approximated by mean-field inference.Source: §3 · p. 3
4. EXPERIMENTS
Experiments compare FCN, adversarial, CRF, and multi-scale variants on INbreast and DDSM-BCRP using Dice and boundary accuracy. The combined model achieves the best reported results, with adversarial training improving segmentation, including around boundaries.
4. EXPERIMENTS
- The compared variants include position-prior FCN, adversarial FCN, end-to-end Joint FCN-CRF, and adversarial FCN-CRF, with multi-scale counterparts using four FCNs.The multi-scale prediction averages the four FCN outputs.Source: §4 · p. 3
- The experiments evaluate INbreast and DDSM-BCRP using Dice, with training and test splits of 58 and 58 mass regions on INbreast and 84 and 87 ROIs on DDSM-BCRP.Each training set was expanded fourfold by horizontal, vertical, and combined flips.Source: §4 · p. 3
- Combining all components achieves the best performance, with 0.97% improvement on INbreast and 1.3% on DDSM-BCRP.The authors report a McNemar’s Chi-Square Test p-value < 0.001 versus model on INbreast.Source: §4 · pp. 3–4, Table 2 · p. 3
- The enhanced FCN improves Dice by 0.25% over traditional FCN on INbreast; adversarial training adds 0.4% on average and structured learning adds 0.3%.The authors report multi-scale modeling as the largest contributor among the examined components.Source: §4 · p. 3
- Visual examples on INbreast show sharper, more accurate segmentation boundaries for adversarial variants than for methods without adversarial training.Rows compare FCN, Adversarial FCN, Joint FCN-CRF, and Adversarial FCN-CRF; columns are test samples.Source: §4 · p. 4, Figure 3 · p. 4
- Adversarial FCN-CRF achieves 2–3% higher trimap accuracy on average than Joint FCN-CRF, and Adversarial FCN also outperforms FCN.Trimaps evaluate accuracy around ground-truth mass boundaries.Source: §4 · p. 4
5. CONCLUSION
The conclusion presents an end-to-end adversarial FCN-CRF with a position prior for mammographic mass segmentation. Results on two public datasets are reported as superior, and adversarial training improves boundary accuracy.
5. CONCLUSION
- The end-to-end adversarial FCN-CRF achieves superior performance on two commonly used public datasets.The network adds a position prior and uses adversarial training to address overfitting with small training data.Source: §5 · p. 4
- Adversarial training improves segmentation accuracy around boundaries in trimaps of widths 1–5 on INbreast and DDSM-BCRP.The comparison includes FCN, Adversarial FCN, Joint FCN-CRF, and Adversarial FCN-CRF.Source: Figure 4 · p. 4