INSTANCE-LEVEL STACKING OF DEEPLABV3+ AND SEGFORMER ENSEMBLES WITH EXPLAINABLE FEATURE FUSION FOR NUCLEUS PHENOTYPE CLASSIFICATION IN IMMUNOHISTOCHEMISTRY IMAGES OF LYMPHOMA

Abstract

Accurate nucleus phenotype classification in immunohistochemistry (IHC) images is essential for biomarker quantification and characterization of the tumor microenvironment, yet automated analysis remains difficult because staining intensity, nuclear morphology, overlapping structures, and tissue heterogeneity vary considerably across specimens. To address these challenges, this study presents an explainable instance-level stacking framework that combines convolutional and transformer-based segmentation models with feature-level meta-learning. DeepLabV3+ with a ResNet-50 encoder and SegFormer-B2 were selected because they capture complementary aspects of tissue organization. Both architectures were trained using five-fold cross-validation, and fold predictions were averaged to obtain ensemble probability maps. Instead of performing direct pixel-level fusion, probability information was aggregated for each annotated nucleus using instance masks. This enabled the extraction of morphological descriptors, probability statistics, and features describing agreement and disagreement between the two models. These complementary descriptors were integrated through an XGBoost stacking classifier. Experiments on the LyNSeC dataset containing 86,997 annotated nuclei showed that the proposed framework outperformed individual models and simpler fusion strategies. The centroid-independent model achieved 95.94% accuracy and a ROC-AUC of 99.24%. SHAP analysis indicated that SegFormer-derived probabilities and inter-model agreement features contributed most strongly to classification, while ablation experiments confirmed the benefit of combining multiple feature categories.