PAGE-Net

PAGE-Net integrates histopathological images and genomic data to predict cancer patient survival and identify prognostic histopathological and genomic factors.

Key Features:

• Multi-Modal Data Integration: Combines pathological images and genomic information for joint analysis of tissue morphology and molecular profiles.
• Biologically Interpretable Model: Uses pathology-, genome-, and demography-specific layers to provide interpretable, biologically informed representations.
• Innovative Pathology-Specific Layers: Employs a patch-wise texture-based convolutional neural network (CNN) with a patch aggregation strategy to extract global survival-discriminative features from histopathological images without manual annotation.
• Genome-Specific Layers: Adapts the pathway-based sparse deep neural network Cox-PASNet for genomic data to capture pathway-level genetic patterns associated with survival.
• Enhanced Predictive Performance: Demonstrated improved concordance index (C-index) on Glioblastoma Multiforme (GBM) using TCGA and TCIA data, achieving C-index 0.702 versus 0.509 for histopathology-only and 0.640 for Cox-PASNet alone.
• Identification of Prognostic Factors: Simultaneously identifies histopathological and genomic features associated with patient outcomes.

Scientific Applications:

• Oncology research: Supports integrative analyses of tissue morphology and molecular alterations to improve understanding of cancer heterogeneity and prognosis.
• Glioblastoma (GBM) survival analysis: Applied to GBM datasets from TCGA and TCIA for benchmarking multi-modal survival prediction and prognostic factor discovery.

Methodology:

Trains deep learning models on paired histopathological images and genomic data using pathology/genome/demography-specific layers; employs a patch-wise texture-based CNN with patch aggregation for image feature extraction; adapts the pathway-based sparse deep neural network Cox-PASNet for genomic modeling; evaluates performance using concordance index (C-index).