A Novel Tile-Efficient AI Framework for Whole-Slide Histopathology Analysis

Abstract

Whole-slide histopathology images contain rich visual information, but their gigapixel scale creates a challenging machine learning problem. Slides from a single patient can produce thousands of tiles, many of which are redundant, noisy, weakly informative, or unrelated to the target label. In weakly supervised learning, these tiles are commonly organized into large patient-level bags. This increases computational cost and can dilute the predictive signal. This work proposes a weakly supervised learning framework for efficient tile selection from whole-slide images. The framework first extracts tile embeddings using pretrained pathology foundation models, then applies dimensionality reduction, clustering, and self-supervised attention-based selection to build compact patient-level bags. These selected bags are evaluated against all-tile baselines using attention-based multiple-instance learning. Experiments are performed using whole-slide images from The Cancer Genome Atlas in a source-to-target setting. Breast cancer cases serve as the labeled source cohort, while ovarian cancer cases serve as an external target cohort. Candidate strategies are ranked using breast cancer validation AUC; the best strategy is then locked and applied to ovarian cancer. The task was to predict the mutation status of TP53. Target-cohort mutation labels are not used during training, strategy selection, checkpoint selection, or threshold selection. A domain-adversarial extension is also evaluated using unlabeled target-cohort bags to reduce domain shift. Across source-cohort candidate configurations, selected-tile bags often matched or outperformed all-tile baselines while retaining far fewer instances per patient. The validation-selected strategies also remained competitive when applied without reranking to the external cohort. These findings suggest that exhaustive tile aggregation is not always required for weakly supervised whole-slide learning. Tissue-composition analysis, stitched-slide visualization, and pseudo-expression profiles further show that selected tiles follow structured visual and image-derived molecular patterns rather than behaving like random slide regions. The approach presented in this thesis provides a framework for preserving or improving predictive performance while reducing computational cost in histopathology image analysis. By reducing the number of tiles processed during patient-level learning, the framework can lower memory demand, training time, and inference cost. With this in mind, replacing exhaustive aggregation with informed tile selection can make weakly supervised computational pathology more efficient, more inspectable, and better suited for large-scale clinical implementation.