GSAL: Geometric structure adversarial learning for robust medical image segmentation

• A new learning mechanism for medical image segmentation. We introduce a novel Geometric Structure Learning Mechanism (GSLM) that enhances model learning "focus, path, and difficulty". It enables geometric structure attention learning to bridge image features with large differences, thus capturing the contextual dependencies of images. The image features maintain consistency and continuity along the internal and external geometry structure, which improves the integrity and boundary accuracy of the segmentation results. To the best of our knowledge, we are the first attempt to explicitly establish the target's geometric structure, which has been successfully applied to medical image segmentation. • A novel geometric structure adversarial learning for robust medical image segmentation. We present the geometric structure adversarial learning model (GSAL) that consists of a geometric structure generator, skeleton-like and boundary discriminators, and a geometric structure fusion sub-network. The generator yields the geometric structure that preserves interior characteristics consistency and external boundary structure continuity. The dual discriminators are trained simultaneously to enhance and correct the characterization of interior structure and boundary structure, respectively. The fusion sub-network aims to fuse the geometric structure that optimized by adversarial learning to refine the final segmentation results with higher credibility. • State-of-art results on widely-used benchmarks. Our GSAL achieves SOTA performance on a variety of benchmarks, including Kvasir&CVC-612 dataset, COVID-19 dataset, and LIDC-IDRI dataset. It confirms the robustness and generalizability of our framework. In addition, our method has great advantages in terms of the integrity and boundary accuracy of the segmentation target compared to other competitive methods. GSAL can also achieve a considerable trade-off in terms of accuracy, inference speed, and model complexity, which helps deploy in clinical practice systems. Automatic medical image segmentation plays a crucial role in clinical diagnosis and treatment. However, it is still a challenging task due to the complex interior characteristics (e.g. , inconsistent intensity, low contrast, texture heterogeneity) and ambiguous external boundary structures. In this paper, we introduce a novel geometric structure learning mechanism (GSLM) to overcome the limitations of existing segmentation models that lack learning "focus, path, and difficulty." The geometric structure in this mechanism is jointly characterized by the skeleton-like structure extracted by the mask distance transform (MDT) and the boundary structure extracted by the mask distance inverse transform (MDIT). Among them, the skeleton-like and boundary pay attention to the trend of interior characteristics consistency and external structure continuity, respectively. With this idea, we design GSAL, a novel end-to-end geometric structure adversarial learning for robust medical image segmentation. GSAL has four components: a geometric structure generator, which yields the geometric structure to learn the most discriminative features that preserve interior characteristics consistency and external boundary structure continuity, skeleton-like and boundary structure discriminators, which enhance and correct the characterization of internal and external geometry to mutually promote the capture of global contextual dependencies, and a geometric structure fusion sub-network, which fuses the two complementary and refined skeleton-like and boundary structures to generate the high-quality segmentation results. The proposed approach has been successfully applied to three different challenging medical image segmentation tasks, including polyp segmentation, COVID-19 lung infection segmentation, and lung nodule segmentation. Extensive experimental results demonstrate that the proposed GSAL achieves favorably against most state-of-the-art methods under different evaluation metrics. The code is available at: https://github.com/DLWK/GSAL. [ BSTRACT FROM AUTHOR] Copyright of Pattern Recognition is the property of Pergamon Press - An Imprint of Elsevier Science and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full . (Copyright applies to all s.)

Learning label diffusion maps for semi-automatic segmentation of lung CT images with COVID-19

[Display omitted] • A fully end-to-end deep learning approach for COVID-19 CT image segmentation. • The trained model induces the diffusion of seeds by taking as input a marked slice. • The method learns diffusion maps by predicting edge weights via deep contour learning, • The use of deep contour learning and seeded segmentation as an integrated method. Deep Learning (DL) has become one of the key approaches for dealing with many challenges in medical imaging, which includes lung segmentation in Computed Tomography (CT). The use of seeded segmentation methods is another effective approach to get accurate partitions from complex CT images, as they give users autonomy, flexibility and easy usability when selecting specific targets for measurement purposes or pharmaceutical interventions. In this paper, we combine the accuracy of deep contour leaning with the versatility of seeded segmentation to yield a semi-automatic framework for segmenting lung CT images from patients affected by COVID-19. More specifically, we design a DL-driven approach that learns label diffusion maps from a contour detection network integrated with a label propagation model, used to diffuse the seeds over the CT images. Moreover, the trained model induces the diffusion of the seeds by only taking as input a marked CT-scan, segmenting hundreds of CT slices in an unsupervised and recursive way. Another important trait of our framework is that it is capable of segmenting lung structures even in the lack of well-defined boundaries and regardless of the level of COVID-19 infection. The accuracy and effectiveness of our learned diffusion model are attested to by both qualitative as well as quantitative comparisons involving several user-steered segmentations methods and eight CT data sets containing different types of lesions caused by COVID-19. [ FROM AUTHOR]