Source-linked AI summary
Breast Tumor Segmentation and Shape Classification in Mammograms using Generative Adversarial and Convolutional Neural Network
Vivek Kumar Singh, Hatem A. Rashwan, Santiago Romani, Farhan Akram, Nidhi Pandey, Md. Mostafa Kamal Sarker, Adel Saleh, Meritexell Arenas, Miguel Arquez, Domenec Puig, Jordina Torrents-Barrena
TL;DR
Breast tumor segmentation and classification remain challenging because mammograms have low signal-to-noise ratios and tumors vary in shape, size, appearance, texture, and location. The paper proposes a two-stage approach using a cGAN for tumor ROI segmentation followed by shape classification, with cGAN-ResNet101 improving Dice and IoU by about 1% to 3% over non-GAN methods.
Problem
Breast tumor segmentation and classification remain challenging because of low signal-to-noise ratios and tumor variability in shape, size, appearance, texture, and location.
Method
The paper proposes a two-stage approach that first segments breast tumor ROIs using a cGAN and then classifies them.
Results
About 1% to 3% improvement in both Dice and IoU metrics was achieved by cGAN-ResNet101 compared with other non-GAN methods.
Takeaways & Limitations
The proposed method achieves similar or better segmentation accuracy compared with other methods.
Takeaways & Limitations
The segmentation model requires the tumor to be fully contained in the ROI; three samples were mis-segmented because they contained two tumors.
Abstract
from arXiv · showhide
Mammogram inspection in search of breast tumors is a tough assignment that radiologists must carry out frequently. Therefore, image analysis methods are needed for the detection and delineation of breast masses, which portray crucial morphological information that will support reliable diagnosis. In this paper, we proposed a conditional Generative Adversarial Network (cGAN) devised to segment a breast mass within a region of interest (ROI) in a mammogram. The generative network learns to recognize the breast mass area and to create the binary mask that outlines the breast mass. In turn, the adversarial network learns to distinguish between real (ground truth) and synthetic segmentations, thus enforcing the generative network to create binary masks as realistic as possible. The cGAN works well even when the number of training samples are limited. Therefore, the proposed method outperforms several state-of-the-art approaches. This hypothesis is corroborated by diverse experiments performed on two datasets, the public INbreast and a private in-house dataset. The proposed segmentation model provides a high Dice coefficient and Intersection over Union (IoU) of 94% and 87%, respectively. In addition, a shape descriptor based on a Convolutional Neural Network (CNN) is proposed to classify the generated masks into four mass shapes: irregular, lobular, oval and round. The proposed shape descriptor was trained on Digital Database for Screening Mammography (DDSM) yielding an overall accuracy of 80%, which outperforms the current state-of-the-art.
1. Introduction
Breast tumor segmentation and classification remain difficult because mammograms have low signal-to-noise ratios and tumors vary widely in morphology and appearance. The paper proposes a two-stage system that segments tumors into binary masks and classifies those masks into four shape types.
- Morphological features such as tumor shape and margin type play crucial roles in breast tumor diagnosis.
- Tumor segmentation and classification remain challenging because mammograms have low signal-to-noise ratios and tumors vary in shape, size, appearance, texture, and location.
- The proposed pipeline first segments a breast tumor as a binary mask and then classifies the mask as irregular, lobular, oval, or round.
- The shape classifier is designed to learn morphological features from binary masks rather than texture, intensity, or edge information.
- The paper adapts a cGAN for breast tumor segmentation, targeting reliable learning when labeled mammogram boundaries are scarce without increasing prediction-time computational complexity.
- Experiments on public and private databases report performance that outperforms current state-of-the-art methods for tumor segmentation and shape classification.
2. Related Work
Prior work applies deep networks, graphical models, contour methods, and adversarial learning to medical-image segmentation and shape classification. Existing breast-mammogram systems combine diverse segmentation and classification strategies, while this paper builds on cGAN-based tumor segmentation and CNN-based shape classification.
- Deep semantic-segmentation methods use encoder-decoder architectures, including FCN, U-Net, and SegNet, to produce dense image predictions.
- Medical-image segmentation research includes multi-scale models, adversarial methods, and transfer learning from networks such as ResNet and VGG pretrained on ImageNet.
- Breast tumor segmentation has used region growing, artificial and cellular neural networks, graphical models, contour fitting, and FCN-CRF combinations.
- Some approaches combine segmented tumor regions with classifiers to distinguish tumor categories or benign and malignant cases.
- The authors previously adapted a cGAN image-to-image translation algorithm for two-dimensional mammogram tumor segmentation and reported state-of-the-art performance on public and private databases.
- Medical shape-classification work includes Fourier descriptors, topological and geometric representations, spectral graph wavelets, point clouds, and CNN-based descriptors.
- Breast-tumor shape studies classify categories including irregular, lobular, oval, and round using CNN or GAN-based approaches, with prior work also using DDSM.
3. Proposed Methodology
The proposed methodology uses a conditional GAN to segment breast tumors from mammogram ROIs, then classifies the generated masks into four shape categories with a CNN.
- Tumor Segmentation Model (cGAN): The cGAN processes mammogram tumor ROIs to generate binary breast-tumor masks.Inputs are prepared at 256×256 pixels and may use full, loose, or tight ROI framing.
- Tumor Segmentation Model (cGAN): The discriminator evaluates whether a mask paired with its ROI resembles a real or synthetic segmentation.Its output is a 30×30 array of local segmentation-likelihood values.
- Tumor Segmentation Model (cGAN): The cGAN conditions both networks on the ROI, fitting segmentation to tumor appearance and supporting learning with few training samples.Generator and discriminator optimization proceeds concurrently.
- Tumor Segmentation Model (cGAN): The generator combines adversarial and Dice losses to encourage both realistic mask outlines and accurate tumor regions.The Dice term fosters rough mask-shape prediction, while the adversarial term fosters accurate border prediction.
- Shape Classification Model (CNN): A multi-class CNN classifies generated binary masks as irregular, lobular, oval, or round.The mask is downsampled to 64×64 pixels before classification, and morphological features are extracted to simplify the problem.
4. Experiments and Discussion
The proposed models were evaluated using two public mammography datasets and one private dataset.
- Experiments and Discussion: The experiments evaluated the proposed models on two public mammography datasets and one private dataset.The supplied passage identifies the evaluation scope but does not provide dataset names or quantitative results.
INbreast dataset2
INbreast is a public mammography database containing 115 cases and 410 mammograms; 106 tumor images with ground-truth masks were used for segmentation testing.
- INbreast dataset: INbreast contains 115 cases and 410 mammograms covering masses, calcifications, asymmetries, and distortions.The database is publicly available.
- INbreast dataset: The segmentation model was tested on 106 breast-tumor images with corresponding ground-truth binary masks.These images were drawn from the INbreast database.
DDSM dataset
DDSM provided 1,168 breast-tumor cases with shape labels for training and testing the tumor-shape classification model.
- DDSM dataset: DDSM is a publicly available digital database for screening mammography containing 2,620 mammography studies.It was used for tumor-shape classification.
- DDSM dataset: The shape-classification set included 1,168 breast tumors with ground-truth labels.Labels comprised 504 irregular, 473 lobular, 115 oval, and 76 round tumors.
- DDSM dataset: The tumor-shape classification model used 75% of images for training and the remainder for testing.The passage describes the split at the image level.
Hospital Sant Joan de Reus dataset
The private dataset contains 300 malignant tumors with radiologist-provided masks, and evaluates segmentation across three ROI cropping strategies against ten alternative models. The proposed method generally achieves the strongest segmentation results, while its shape classifier reaches 80% overall accuracy on DDSM and the segmentation stage has explicit scope limitations.
- Dataset and evaluation: 300 malignant tumors, with radiologist-provided binary masks, comprise the private dataset; 220 images train the cGAN and 80 test it.The tumors include Luminal-A, Luminal-B, Her-2 and Basal-like samples.
- Dataset and evaluation: Three ROI crops—full mammogram, loose frame and tight frame—are evaluated using Dice and IoU against ten segmentation models.The loose frame balances foreground and background pixels, while the tight frame emphasizes tumor pixels.
- Segmentation results: 89.99% Dice and 81.81% IoU are obtained by the proposed model on private-dataset loose frames, nearly 9% and 10% above the second-best model, respectively.SegNet-VGG16 is the second-best Dice model at 81.09%, while cGAN-ResNet101 is the second-best IoU model at 71.91%.
- Shape classification: 80% overall accuracy is achieved for four tumor-shape classes, with 84% accuracy for oval and 89% for round tumors.Irregular and lobular shapes are mostly assigned to Luminal-A and Luminal-B samples, whereas oval and round shapes indicate Her-2 and Basal-like samples.
- Limitations: The segmentation model requires a prior tumor location and fails when tumors are not fully contained in the ROI; shape classification relies only on DDSM because it is the only public dataset with shape labels.Three samples with a second tumor partly outside the crop are mis-segmented, and more databases are required to improve four-class accuracy.
5. Conclusion
The paper presents a two-stage framework that segments breast tumors with a cGAN and classifies their shapes from generated masks using a CNN. Experiments report strong segmentation and shape-classification performance, while also identifying crop requirements and future extensions.
- 5. Conclusion: The proposed two-stage method first segments a breast tumor ROI using a cGAN, then classifies its shape from the resulting binary mask.The CNN-based shape descriptor focuses on the tumor’s morphological structure rather than the original mammogram’s pixel variations.
- 5. Conclusion: 1% to 3% improvement in both Dice and IoU is achieved by cGAN-ResNet101 compared with non-GAN methods.The results support a role for the adversarial network in optimizing the generative network.
- 5. Conclusion: 10% improvement over cGAN-ResNet101 is obtained by training the proposed model from scratch and replacing L1 loss with Dice loss on loose frame crops.This comparison concerns the stated training and loss configuration on the given datasets.
- 5. Conclusion: Similar or better segmentation accuracy is obtained with a tight frame crop, whereas full-mammogram segmentation yields low accuracy for all models.The method therefore depends on the crop setting used around the tumor.
- 5. Conclusion: 80% overall accuracy is achieved for classifying tumor shapes, with the CNN inferring shape from the cGAN-generated binary mask.The reported shape inference uses morphological information, including irregular, lobular, oval, and round shapes.
- 5. Conclusion: Tumor shape shows an observed association with molecular subtypes: Luminal-A and -B samples are mostly irregular, while Her-2 and Basal-like samples are mostly regular.The paper indicates that tumor shape can be considered when inferring molecular subtype.