fig1

Figure 1. The PhenoSSP framework for deep spatial feature extraction. (A) The End-to-End Pipeline. Raw 7-color mIF images undergo cell segmentation (Mesmer) to generate single-cell patches. PhenoSSP extracts deep features to construct a final annotated spatial map, enabling downstream spatial neighborhood analysis; (B) The Hierarchical Classification Strategy. To address severe class imbalance in tissue samples, a “Coarse Classifier” first separates major lineages (Epithelial, Immune, Other), filtering out non-immune noise before an “Immune Expert Classifier” refines specific immune subtypes; (C) Model Architecture. The backbone features a ViT-S/16 initialized with DINOv2 weights. Key innovations include Marker Embeddings to handle multi-channel inputs and an MLP-based Classifier Head that processes the global [CLS] token from 64 × 64 pixel patches for precise phenotype prediction. mIF: Multiplex immunofluorescence; MLP: multi-layer perceptron; ViT: Vision Transformer.