Unsupervised domain adaptation for the segmentation of breast tissue in mammography images.

BACKGROUND AND OBJECTIVE: Breast density refers to the proportion of glandular and fatty tissue in the breast and is recognized as a useful factor assessing breast cancer risk. Moreover, the segmentation of the high-density glandular tissue from mammograms can assist medical professionals visualizing and localizing areas that may require additional attention. Developing robust methods to segment breast tissues is challenging due to the variations in mammographic acquisition systems and protocols. Deep learning methods are effective in medical image segmentation but they often require large quantities of labelled data. Unsupervised domain adaptation is an area of research that employs unlabelled data to improve model performance on variations of samples derived from different sources. METHODS: First, a U-Net architecture was used to perform segmentation of the fatty and glandular tissues with labelled data from a single acquisition device. Then, adversarial-based unsupervised domain adaptation methods were used to incorporate single unlabelled target domains, consisting of images from a different machine, into the training. Finally, the domain adaptation model was extended to include multiple unlabelled target domains by combining a reconstruction task with adversarial training. RESULTS: The adversarial training was found to improve the generalization of the initial model on new domain data, demonstrating clearly improved segmentation of the breast tissues. For training with multiple unlabelled domains, combining a reconstruction task with adversarial training improved the stability of the training and yielded adequate segmentation results across all domains with a single model. CONCLUSIONS: Results demonstrated the potential for adversarial-based domain adaptation with U-Net architectures for segmentation of breast tissue in mammograms coming from several devices and demonstrated that domain-adapted models could achieve a similar agreement with manual segmentations. It has also been found that combining adversarial and reconstruction-based methods can provide a simple and effective solution for training with multiple unlabelled target domains.

3D Pulmonary Artery Segmentation from CTA Scans Using Deep Learning with Realistic Data Augmentation.

The characterization of the vasculature in the mediastinum, more specifically the pulmonary artery, is of vital importance for the evaluation of several pulmonary vascular diseases. Thus, the goal of this study is to automatically segment the pulmonary artery (PA) from computed tomography angiography images, which opens up the opportunity for more complex analysis of the evolution of the PA geometry in health and disease and can be used in complex fluid mechanics models or individualized medicine. For that purpose, a new 3D convolutional neural network architecture is proposed, which is trained on images coming from different patient cohorts. The network makes use a strong data augmentation paradigm based on realistic deformations generated by applying principal component analysis to the deformation fields obtained from the affine registration of several datasets. The network is validated on 91 datasets by comparing the automatic segmentations with semi-automatically delineated ground truths in terms of mean Dice and Jaccard coefficients and mean distance between surfaces, which yields values of 0.89, 0.80 and 1.25 mm, respectively. Finally, a comparison against a Unet architecture is also included.