Contextual Classification Of Tumor Growth Patterns In Digital Histology Slides

Patch-based image classification approaches are frequently employed in digital pathology applications with the goal to automatically delineate diagnostically important regions in digital slides. However, patches into which a slide is partitioned are often classified singly and sequentially without additional context. To address this issue, we tested a contextual classification of image patches with soft voting applied to a multi-class classification problem. The context comprised five or nine overlapping patches. The classification is performed using convolutional neural networks (CNNs) trained to recognize four histologically distinct growth patterns of lung adenocarcioma and non-tumor areas. Classification with soft voting outperformed sequential classification of patches yielding higher whole slide classification accuracy. The F1-scores for the four tumor growth patterns improved by 3% and 4.9% when the context consisted of five and nine neighboring patches, respectively. We conclude that the context can improve classification performance of areas sharing the same histological features. Soft voting is a non-trainable approach and therefore straightforward to implement. However, it is computationally more expensive than the classical single patch-based approach.