HEMnet

HEMnet leverages immunohistochemistry molecular labels and deep learning on Haematoxylin and Eosin (H&E) stained tissue images to classify cancerous versus normal tissue and estimate tumor purity.

Key Features:

• Molecular Labeling Integration: Uses immunohistochemistry (IHC) to provide initial molecular labels for cancer cells within Haematoxylin and Eosin (H&E) stained tissue images.
• Two-Step Deep Learning Process: Employs a two-step approach where molecular labels are transferred to overlapping clinical histopathological images and used to train a deep learning cancer classifier.
• High Accuracy in Tumor Identification: Demonstrated capability to distinguish colorectal cancer from normal tissue and to estimate tumor purity, with validation using histopathology images from The Cancer Genome Atlas (TCGA).
• Automation and Transferability: Implements a molecular transfer method to enable automated delineation of various tumor types when a cancer-oriented molecular stain is available for training.

Scientific Applications:

• Pathological annotation: Reduces reliance on manual histopathologic annotation by deriving labels from molecular staining.
• Cancer diagnosis and tumor purity estimation: Supports classification of colorectal cancer versus normal tissue and provides tumor purity estimates from H&E images.
• Tumor delineation and cross-tumor transfer: Facilitates automated delineation of tumors and application of molecularly trained classifiers across different cancer types given appropriate molecular stains.

Methodology:

Integrates immunohistochemistry for molecular labeling of H&E images, then applies a deep learning-based two-step classification where molecular labels are used to train a classifier on overlapping histopathological images.