73. Pathomics-based machine mearning versus deep learning: Which is a better approach for whole slide image analyses?

Pathology relies on examining H&E-stained FFPE tissue sections via microscopy for diagnosis. Pathomics quantifies features from digitized FFPE images, or whole slide images (WSIs), reflecting tissue and cellular structures potentially linked to gene expression patterns seen in RNA sequencing. Although Deep Learning (DL) methods have advanced gene expression prediction from WSIs, understanding the pathomic features affecting model predictions is challenging.

Our study analyzed 89 FFPE breast cancer tissue slide images from the TCGA registry, extracting around 300 pathomic features using HistomicsTK. These features, representing cell morphometry, intensity, and gradient, were assessed within tumor regions annotated by pathologists. We selected the most heterogeneous and correlating features for a multitask regression model, predicting gene expression levels with high accuracy (AUC > 0.8) for two biomarkers, MFAP5 and MXRA8.

In contrast, the ResNet-50 DL model trained on random WSI patches showed lower AUC scores for these biomarkers and did not interpret pathomic features that contribute to gene expression predictions. Literature suggests MFAP5 upregulation in breast carcinomas correlates with poor prognosis, while MXRA8 modulates triple-negative breast cancer progression.

The study concludes that Pathomics-based Machine Learning outperforms DL in predicting gene expression from FFPE WSIs in invasive breast carcinoma, providing a more effective tool for understanding the disease at the molecular level.