A Review of Advanced Deep Learning Methods of Multi-Target Segmentation for Breast Cancer WSIs

Abstract

Breast cancer is one of the most common cancers among women, with its heterogeneity posing significant challenges for diagnosis and treatment, profoundly impacting patient prognosis and quality of life. Whole Slide Imaging (WSI) in digital pathology provides high-resolution images that enable a comprehensive examination of the tumor microenvironment, offering advanced tools for breast cancer diagnosis and prognostic evaluation. However, manually reviewing whole slide images (WSIs) for tissue segmentation is time-consuming and prone to errors, highlighting the need for multi-target deep learning models to automate the segmentation of these complex structures. Multi-target segmentation offers distinct advantages by simultaneously processing multiple interrelated tissue regions within a single image, thereby enhancing accuracy and efficiency. Despite the potential of deep learning techniques in automating pathological analysis, their clinical adoption faces significant challenges. To address these, this paper proposes six criteria focused on clinical acceptability of deep learning methods: inherent limitations of WSIs, feature extraction, annotation requirements, efficiency, automated quantification, and interpretability. A rigorous review of publicly available datasets and deep learning methods identifies key challenges for clinical adoption. Following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines, this review analyzes 29 core articles, highlighting the critical role of multi-target segmentation in breast cancer digital pathology while assessing the limitations of these techniques in clinical applications. Based on this analysis, this paper proposes six criteria to enhance the diagnostic performance of deep learning methods in multi-target segmentation for breast cancer digital pathology and to improve the clinical acceptability of deep learning methods.