Abstract
Follow-up studies with RNA extraction from various areas of AI-predicted molecular heterogeneity may improve molecular classifications and thereby AI algorithms trained on these classifications.
Introduction
Muscle-invasive urothelial carcinoma (MIUC) remains a significant clinical challenge, with a 5-year survival rate of approximately 50%. While molecular subtyping (e.g., luminal vs. basal) offers a path toward personalized therapy, its adoption is hindered by the high cost and complexity of RNA sequencing. Furthermore, tumors often exhibit significant intratumoral heterogeneity, meaning a single bulk-sequencing result may not capture the full biological landscape. This study explores an AI-based approach to predict these molecular subtypes directly from standard hematoxylin and eosin (H&E) histology slides.
Methods
The researchers developed a deep learning workflow using the ResNet-18 architecture. Key methodological innovations included:
- Regression-Based Training: Instead of traditional discrete classification, the model was trained to predict Pearson correlation coefficients for each subtype. This approach allows the model to quantify the degree to which a tissue sample resembles a specific molecular profile, better reflecting biological reality.
- Preprocessing: Pathologists annotated tumor areas using QuPath, and slides underwent Macenko stain normalization to ensure consistency across different laboratory settings.
- Explainability: Grad-CAM (Gradient-weighted Class Activation Mapping) was utilized to verify that the model's predictions were based on relevant histological features, such as squamous differentiation in basal/squamous subtypes.
Datasets and Validation
The model was trained and internally tested on the TCGA-BLCA dataset ($N=292$). To ensure robustness, external validation was performed on an independent cohort from the Senckenberg Institute of Pathology (SIP-BLCA, $N=27$).
Key Results and Insights
The regression-based approach significantly outperformed traditional classification, achieving an average AUROC of 0.80 in internal validation and 0.75 in external validation. The model was particularly effective at identifying Luminal and Basal/Squamous subtypes.
A critical finding was the model's ability to map intratumoral heterogeneity. By generating spatial "heatmaps" across whole-slide images, the AI identified co-existing regions of different subtypes within the same tumor. This spatial mapping highlights the limitations of single-site biopsies and bulk sequencing.
Discussion and Future Directions
While the results are promising, the "Stroma-rich" subtype remains difficult to predict due to the complex interplay between tumor and non-tumor cells. Future work will focus on "micro-sequencing"—using AI-generated maps to guide targeted RNA extraction from specific regions. This will allow for direct validation of the AI's heterogeneity predictions and further refinement of the classification algorithms. Ultimately, this technology could provide clinicians with an immediate, low-cost molecular map to guide personalized treatment strategies.
Summary updated