[Paper Review] Conditional Generative Adversarial and Convolutional Networks for X-ray Breast Mass Segmentation and Shape Classification

Vivek Kumar Singh, Santiago Romaní+10 more|arXiv (Cornell University)|May 25, 2018

AI in cancer detection14 references5 citations

TL;DR

This paper proposes a conditional Generative Adversarial Network (cGAN) with U-Net and autoencoder architectures for accurate X-ray breast mass segmentation and subsequent shape classification. The cGAN framework improves segmentation performance by enforcing adversarial consistency with ground truth, achieving over 94% Dice and 89% Jaccard on public and private datasets, while a separate CNN classifies segmented masses into four shape types with 72% accuracy.

ABSTRACT

This paper proposes a novel approach based on conditional Generative Adversarial Networks (cGAN) for breast mass segmentation in mammography. We hypothesized that the cGAN structure is well-suited to accurately outline the mass area, especially when the training data is limited. The generative network learns intrinsic features of tumors while the adversarial network enforces segmentations to be similar to the ground truth. Experiments performed on dozens of malignant tumors extracted from the public DDSM dataset and from our in-house private dataset confirm our hypothesis with very high Dice coefficient and Jaccard index (>94% and >89%, respectively) outperforming the scores obtained by other state-of-the-art approaches. Furthermore, in order to detect portray significant morphological features of the segmented tumor, a specific Convolutional Neural Network (CNN) have also been designed for classifying the segmented tumor areas into four types (irregular, lobular, oval and round), which provides an overall accuracy about 72% with the DDSM dataset.

Motivation & Objective

• To address the challenge of accurate breast mass segmentation in low-contrast, noisy mammograms with limited training data.
• To improve segmentation robustness by leveraging conditional GANs that enforce realism and boundary fidelity through adversarial training.
• To classify segmented tumor regions into four morphological shape types (irregular, lobular, oval, round) for potential correlation with molecular subtypes.
• To evaluate the performance of cGAN-based segmentation on both public (DDSM) and private in-house datasets, comparing with FCN, U-Net, and CRFCNN.
• To explore the correlation between tumor shape and molecular subtypes (Luminal-A/B, HER-2, Basal-like) in a private clinical dataset.

Proposed method

• A cGAN framework is used where the generator learns to produce binary masks of breast masses conditioned on input mammogram ROI crops.
• The generator network is implemented as a U-Net or autoencoder architecture, combining encoder feature extraction with skip connections or no skip connections.
• The discriminator network compares generated masks with ground truth masks to enforce realistic and precise segmentation boundaries.
• Adversarial loss is combined with binary cross-entropy and L1 loss to optimize the generator, improving segmentation fidelity.
• Post-processing morphological filtering (erosion and dilation) is applied to remove small artifacts in the generated masks.
• A separate, lightweight CNN is trained on binary masks to classify tumor shapes into four categories using 10-fold cross-validation.

Experimental results

Research questions

• RQ1 Can a conditional GAN architecture improve breast mass segmentation accuracy, especially with limited training data?
• RQ2 How does the inclusion of skip connections (U-Net vs. autoencoder) in the cGAN generator affect segmentation performance on public and private datasets?
• RQ3 Does the proposed cGAN-based segmentation pipeline outperform state-of-the-art models like FCN, U-Net, and CRFCNN in terms of Dice and Jaccard scores?
• RQ4 Can a simple CNN effectively classify tumor shapes from binary masks with high accuracy?
• RQ5 Is there a statistically significant correlation between tumor shape (irregular, lobular, oval, round) and molecular subtypes (Luminal-A/B, HER-2, Basal-like) in clinical data?

Key findings

• The cGAN-Unet model achieved a Dice coefficient of 95.55% and Jaccard index of 86.48% on the DDSM test set, outperforming FCN, U-Net, and CRFCNN.
• On the private in-house dataset, cGAN-AutoEnc achieved a Dice score of 94.81% and Jaccard of 80.08%, with a sensitivity of 97.26%, indicating superior tumor area detection.
• The cGAN models improved segmentation performance by +6% in Dice and +9% in Jaccard compared to baseline FCN on the DDSM dataset.
• The shape classification CNN achieved an overall accuracy of 72% on the DDSM dataset using 10-fold cross-validation.
• In the private dataset, Luminal-A and Luminal-B subtypes were predominantly associated with irregular and lobular shapes, while oval and round shapes were more common in HER-2 and Basal-like subtypes.
• The cGAN-AutoEnc model showed better generalization on the private dataset than cGAN-Unet, suggesting that simpler architectures may learn more robust tumor features when data is limited.