Document Type

Thesis

Degree Name

Master of Applied Computing

Department

Physics and Computer Science

Program Name/Specialization

Applied Computing

Faculty/School

Faculty of Science

First Advisor

Azam Asilian Bidgoli

Advisor Role

Thesis Supervisor

Abstract

Computational pathology increasingly relies on the analysis of Whole-Slide Images (WSIs), which capture tissue specimens at gigapixel resolution. Because a single slide is far too large to process directly, the

prevailing paradigm decomposes each WSI into thousands of small patches and encodes them as high- dimensional feature embeddings using deep learning backbones. While effective, this paradigm carries a

substantial cost: the resulting collections of patch embeddings are computationally expensive to store and process, and they are frequently dominated by redundant, homogeneous, or otherwise uninformative tissue regions that dilute the diagnostic signal. Existing patch selection methods largely depend on heuristic or representativeness-based sampling and do not explicitly model the inherent tension between the compactness of the retained representation and the diagnostic accuracy it can support. To address this gap, this thesis introduces EvoPS (Evolutionary Patch Selection), a framework that reformulates patch selection within the training embedding space as a multi-objective optimization problem.

Rather than applying a fixed sampling rule, EvoPS employs an evolutionary search based on the Non- dominated Sorting Genetic Algorithm II (NSGA-II) algorithm to simultaneously minimize the number of

selected patch embeddings and maximize the performance of a downstream similarity-search classification task. Crucially, the search operates entirely on pre-computed embeddings and at the level of the training set, while every validation and test slide retains its full complement of patches, preserving a strict and unbiased evaluation protocol. Constraint-preserving genetic operators ensure that every whole-slide image remains represented throughout the search. Rather than returning a single solution, EvoPS produces a Pareto front of optimal trade-off solutions, allowing a user to select the operating point that best balances computational cost against diagnostic performance. By isolating a compact and diagnostically informative subset of training patches, the framework yields higher-quality training representations that reduce memory requirements and limit the extent to which uninformative tissue dilutes the diagnostic content of the aggregated representation. The framework was validated across four major cancer cohorts from The Cancer Genome Atlas (TCGA),

comprising 2,587 whole-slide images, using five histopathology backbones spanning both task-specific con- volutional networks and large-scale foundation models. The results demonstrate that EvoPS can reduce the

required number of training patches by more than 90% and in several settings by over 98%, while consistently maintaining or even improving the final classification F1-score relative to a state-of-the-art patch selection baseline. Complementary analyses reinforce these findings: confusion-matrix comparisons show that EvoPS sharpens the decision boundaries between histologically similar sub-types, feature-space visualizations reveal denser and more separable clusters, and qualitative inspection confirms that the selected patches localize to diagnostically relevant tumor-bearing regions while excluding background tissue. The evolutionary search itself is shown to be lightweight, running within minutes on a standard workstation. Finally, the framework is extended to the inference stage, where a learned selector, trained only to identify EvoPS-selected patches on the training set, without access to any test labels is used to retain a compact subset of patches from each test slide. Across all cohorts and backbones, this test-time selection is shown to preserve classification accuracy, with no statistically significant difference from using every test patch, while discarding a large fraction of patches per slide, demonstrating that substantial patch reduction can be achieved at inference

time without compromising diagnostic performance. Collectively, these results establish EvoPS as a prin- cipled method for creating efficient, accurate, and interpretable WSI representations, empowering users to

select an optimal balance between computational cost and diagnostic performance.

Convocation Year

2026

Convocation Season

Fall

Download

Share

COinS