SWORD Deposit
Permanent URI for this community
ETDs in this community are being checked for completeness and are in the process of being transferred to their respective collections under the FGS ETD collection umbrella.
Browse
Browsing SWORD Deposit by Author "Ali Sadeghi-Naini"
Now showing 1 - 1 of 1
Results Per Page
Sort Options
Item Open Access Deep Learning-Based Detection, Annotation and Staging of Breast Cancer Metastasis on Whole-Slide Histopathology Images of Lymph Nodes(2025-07-23) Tauqeer, Abdullah; Ali Sadeghi-NainiBreast cancer is one of the most frequently diagnosed malignancies and a leading cause of cancer-related mortality among women worldwide. Accurate detection and characterization of breast cancer lymph node metastasis are crucial for effective patient management. In this thesis, two novel deep learning frameworks have been developed and investigated to automate the analysis of hematoxylin and eosin (H&E) stained digital pathology whole-slide images (WSIs) of lymph nodes to streamline the breast cancer pathology workflow in the clinic. First, a nuclei segmentation and classification model, TexSegNet, is introduced to extract detailed cellular-level information from the digital pathology images of biopsied lymph nodes. This model has been trained and evaluated using 7,904 histology images from the PanNuke dataset, which includes 205,343 annotated nuclei across 19 tissue types (2,351 images from breast tissue). TexSegNet leverages a hybrid encoder-decoder architecture, integrating multi-scale convolutions, nuclear texture extraction blocks, advanced attention mechanisms, and a feedback-driven classification branch. Trained on all tissue types of the PanNuke dataset and subsequently fine-tuned on its breast subset, TexSegNet has achieved an overall accuracy of 81.4±0.4% in detecting and classifying nuclei on the breast test set, substantially outperforming established benchmarks such as HoVer-Net, CellViT, and StarDist. Notably, TexSegNet has maintained consistently good performance across various cell types, including underrepresented ones, with F1-scores of 89.3±0.4%, 91.1±0.5%, 88.9±0.8%, and 84.3±0.3% in detecting and classifying neoplastic, epithelial, inflammatory, and connective cell nuclei, respectively. Second, a selective neighborhood attention-based multiple instance learning framework (MIL) is proposed for automated detection, localization, and staging of lymph node metastases in breast cancer. The proposed framework leverages a dual-path feature extractor, incorporating both nuclei segmentation/classification outputs and transformer-based tissue features, alongside a dynamic attention mechanism that selects and emphasizes neighboring patches based on similarity to the target patch. Trained and optimized on 269 WSIs (over 14 million image patches) from the CAMELYON16 dataset and evaluated on an independent test set of 129 WSIs, the model achieves 96.2±1.5% sensitivity, 95.3±2.4% precision, and 95.7±3.1% F1-score in patch-level classification, along with the area under the receiver operating characteristic curve (AUC) of 0.96±0.01 for slide-level classification. External validation on an out-of-distribution (OOD) subset of 30 annotated WSIs from the CAMELYON17 dataset (representing multiple institutions) demonstrates robust generalizability, with a patch-level F1-score of 87.0±1.8% and a slide-level AUC of 0.88±0.03. Furthermore, for 500 WSIs from 100 patients in the CAMELYON17 dataset, the framework demonstrates an excellent performance in patient-level pN-staging, with a quadratic-weighted Cohen’s kappa of 0.94±0.02, indicating near-expert concordance in detecting and classifying the extent of nodal metastasis. Ablation analyses underscore the importance of incorporating nuclei-based features and selective neighborhood attention, with noticeable performance degradation observed when either element is removed. By integrating insights into histologic type, morphology and spatial heterogeneity at the cellular level with contextual information at the tissue level, the proposed framework effectively replicates key aspects of human pathological assessment and shows promise as a decision-support tool for metastatic breast cancer detection, annotation, and staging in the era of digital pathology. Together, these two frameworks, TexSegNet for detailed nuclei segmentation and classification, and the selective neighborhood attention-based MIL model for targeting metastatic tumor detection and staging, form an integrated pipeline for breast cancer pathology assessment on standard H&E-stained slides. Together, they demonstrate strong potential to reduce pathologist workload, reduce subjectivity in diagnosis, and serve as processing tools in digital pathology to derive quantitative biomarkers for further analysis or prognostication. Future work will focus on expanding dataset diversity, enhancing model explainability, and integrating these algorithms into clinical workflows to support precision oncology care.