Abstract
BACKGROUND: Deep learning (DL) has set new standards in cancer diagnosis, significantly enhancing the accuracy of automated classification of whole slide images (WSIs) derived from biopsied tissue samples. To enable DL models to process these large images, WSIs are typically divided into thousands of smaller tiles, each containing 10-50 cells. Multiple Instance Learning (MIL) is a commonly used approach, where WSIs are treated as bags comprising numerous tiles (instances) and only bag-level labels are provided during training. The model learns from these broad labels to extract more detailed, instance-level insights. However, biopsied sections often exhibit high intra- and inter-phenotypic heterogeneity, presenting a significant challenge for classification. To address this, many graph-based methods have been proposed, where each WSI is represented as a graph with tiles as nodes and edges defined by specific spatial relationships. RESULTS: In this study, we investigate how different graph configurations, varying in connectivity and neighborhood structure, affect the performance of MIL models. We developed a novel pipeline, K-MIL, to evaluate the impact of contextual information on cell classification performance. By incorporating neighboring tiles into the analysis, we examined whether contextual information improves or impairs the network's ability to identify patterns and features critical for accurate classification. Our experiments were conducted on two datasets: COLON cancer and UCSB datasets. CONCLUSIONS: Our results indicate that while incorporating more spatial context information generally improves model accuracy at both the bag and tile levels, the improvement at the tile level is not linear. In some instances, increasing spatial context leads to misclassification, suggesting that more context is not always beneficial. This finding highlights the need for careful consideration when incorporating spatial context information in digital pathology classification tasks.