Abstract 1160: Multi-scale deep learning feature distributions distinguish invasive ductal and lobular breast cancer slides

Abstract

Abstract Advances in digital pathology have streamlined use of deep learning methodologies to connect morphology and phenotype in hematoxylin and eosin(H&E) stained images. Whole slide images (WSIs) are gigantic, necessitating a tile-based analysis: a WSI is broken into smaller images called tiles, each tile is analyzed separately, and tile level information are then combined to get a slide level prediction. Context aware models integrating tiles of different sizes have been used to combine morphological features present at different scales, and attention-based pipelines have successfully been used to combine tile level information. Pipelines that combine attention-based and context aware learning are less studied. While both methodologies have shown utility when used individually, models that combine both tend to suffer from a large number of parameters to estimate, which frustrates reliable training. Here we show percentiles of deep learning features extracted at different scales serve as a crude multiscale attention mechanism. The proposed model separates TCGA-BRCA invasive ductal and lobular carcinoma FFPE WSIs (AUC=0.85±0.02), which improves upon the performance of the fully tile-based model of [1] (AUC=0.80 ±0.04). The proposed approach enjoys higher average AUC and lower standard deviation across multiple random train/test splits of data. Our results suggest percentile-based feature construction is an interesting alternative when reliable training of fully deep learning models is challenging.