University of British Columbia
Mixture of multi-scale prototypes learning for anomaly detection in digital pathology
Abstract
dc:descriptionDeep learning models for medical image analysis hold transformative potential for clinical practice, yet their widespread adoption remains constrained by a fundamental limitation: they provide no indication when a slide falls outside their normative training distribution. This failure mode is acutely consequential for rare cancer subtypes, which occur too infrequently to populate supervised training sets but whose misclassification carries direct consequences for patient survival and treatment assignment. This thesis introduces Mixture Modeling on Scaled Attributes model, a multi-scale multiple instance learning framework for anomaly detection in whole slide images (WSIs). The model addresses two principal limitations of existing approaches. First, prior prototypical anomaly detection methods model each class as a unimodal distribution, which is inadequate for biologically heterogeneous cancer tissue. The model replaces this with a mixture of K probabilistic distributions per class, capturing intra-class morphological variability and establishing low-probability regions that anomalous samples may occupy. Second, most digital pathology anomaly detection systems operate at a single fixed magnification, discarding complementary diagnostic information available at other scales. This model integrates a Spatially-aware Gated Attention module positioned between consecutive magnifications that simultaneously learns an optimal WSI-level representation and a differentiable patch selection policy for higher-resolution refinement, mirroring the multi-scale diagnostic workflow of expert pathologists. The model is evaluated on four large, multi-institutional datasets encompassing both near- and far-anomaly detection scenarios under in-distribution and domain-shift conditions. The model achieves the highest or second-highest performance on all five test sets on at least one metric per dataset, with statistically significant improvements over the strongest competing method in every case (paired bootstrap, p < 0.05). Prototype assignment and attention visualization confirm that learned components correspond to interpretable morphological tissue sub-structures, supporting clinical transparency. These results demonstrate that mixture-based regularization and end-to-end differentiable multi-scale learning produce reliable uncertainty awareness and facilitate rare cancer subtype identification, representing a meaningful step toward safer and more effective integration of artificial intelligence into pathology workflows.
Degree
thesis:*- Name thesis:degree_name
- Master of Science - MSc
- Level thesis:degree_level
- master's
- Discipline thesis:degree_discipline
- Bioinformatics
- Grantor dc:publisher
- University of British Columbia
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Shi, Xin Ping
- Contributors dc:contributor
-
- Bashashati, Ali
Rights
dc:rights- Statement dc:rights
-
- Attribution-NonCommercial-NoDerivatives 4.0 International
- Language dc:language
- eng
Identifiers
dc:identifier.*- Handle dc:identifier
- http://hdl.handle.net/2429/95490