SuperHistopath

SuperHistopath maps global context features in whole-slide digital histopathology images using deep learning to segment and classify histological tissue components and quantify morphological heterogeneity of tumors.

Key Features:

• Global context mapping: Uses deep learning to capture intricate morphological heterogeneity of tumors from whole-slide digital histopathology images.
• Computational efficiency: Classifies a whole-slide image in approximately 5 minutes and requires about 30 minutes to train the network.
• Segmentation (SLIC superpixels): Employs the linear iterative clustering (SLIC) superpixels algorithm at 5x magnification to form homogeneous spatial units that adhere to tissue region boundaries without requiring boundary delineation for ground-truth annotation.
• CNN classification: Applies a convolutional neural network to classify superpixels into tumor tissue, stroma, necrosis, lymphocyte clusters, differentiating regions, fat, hemorrhage, and normal tissue.

Scientific Applications:

• Classification accuracy: Achieved accuracies of 98.8% for melanomas, 93.1% for triple-negative breast cancers, and 98.3% for transgenic mouse models of high-risk childhood neuroblastoma.
• Phenotypic discovery: Enabled discovery of significant phenotypic differences in neuroblastoma transgenic mouse models that emulate genomic variants associated with high-risk disease.
• Patient stratification: Identified prognostic markers in melanoma, including a high lymphocyte-to-tumor superpixel ratio associated with favorable prognosis (p = 0.015) and a low stroma-to-tumor ratio associated with favorable prognosis (p = 0.028).

Methodology:

Computational methods explicitly include linear iterative clustering (SLIC) superpixel segmentation at 5x magnification, deep learning-based global context feature mapping, convolutional neural network classification of superpixels, and reported whole-slide classification (~5 minutes) and network training (~30 minutes) times.