Deep learning-based spatial analysis on tumor and immune cells of pathology images predicts MIBC prognosis

Muscle-invasive bladder cancer (MIBC) is a highly aggressive disease with a poor prognosis. This study aims to explore the correlation between the spatial distribution of lymphocyte aggregates and the prognosis of MIBC using deep learning.

Whole-slide images (WSIs) and clinical information of MIBC patients were collected from The Cancer Genome Atlas (TCGA) and the HBlaU079Su01 microarray. After manually annotating tumor and lymphocyte aggregation regions, we developed an automated analysis pipeline, including segmentation and quality control of patches. A convolutional neural network (CNN) classification model was constructed. Based on the definition of the border region of tumor cell nests, we assessed 12 spatial indicators for different patch types within, around and outside the tumor cluster. Using multiomics, we explored the relationships among spatial indicators, clinical factors, the immune microenvironment and molecular mechanisms.

We obtained 301 MIBC WSIs from the TCGA and 43 MIBC WSIs from the HBlaU079Su01 microarray. Two independent CNN models showed excellent classification performance in identifying patches containing tumor cells (AUC = 0.944, 95% CI: 0.942-0.945) and lymphoid aggregates (AUC = 0.934, 95% CI: 0.932-0.937) were constructed. Grad-CAM analysis revealed that some patches contained both tumor cells and lymphocytes. The prognostic value of Lymph_inside % was validated in the HBlaU079Su01 microarray. Multiomics analysis demonstrated that Lymph_inside % demonstrated significant positive correlations with antitumor immune cell abundance and the apoptosis pathway.

Lymph_inside % can be an effective biomarker for predicting MIBC prognosis. This study suggests a novel approach for the development of new prognostic biomarkers based on the spatial distribution of lymphocyte aggregates.