University of British Columbia
Anatomy-aware hierarchical multiple instance learning for interpretable chronic obstructive pulmonary disease diagnosis from chest CT
Abstract
dc:descriptionChronic obstructive pulmonary disease (COPD) is a progressive respiratory disorder characterized by persistent airflow limitation and heterogeneous structural lung changes, yet it remains substantially underdiagnosed. Spirometry, the diagnostic reference standard, provides only a global measure of airflow obstruction and does not capture the spatial distribution of disease. Chest computed tomography (CT) directly visualises COPD-related abnormalities and is increasingly available through lung cancer screening programmes; however, manual assessment is challenging at scale. Multiple instance learning (MIL) is well suited to this setting, as diagnostic labels are assigned at the patient level while abnormalities are distributed heterogeneously across CT images. Existing MIL approaches, however, typically treat slices as an unordered set and provide limited anatomical interpretability. This thesis presents AnatoMIL, an anatomy-aware hierarchical MIL framework for COPD detection, severity estimation, and regional phenotype analysis from chest CT. The proposed framework aggregates slice embeddings into lobe-level representations using soft lobe-fraction assignments derived from automated segmentation, followed by gated attention-based patient-level aggregation. This design produces a lobe contribution profile alongside each prediction while preserving weakly supervised learning from patient-level labels. A dual-head objective jointly performs binary COPD detection and ordinal GOLD severity estimation. The framework was developed on the Pan-Canadian Early Detection of Lung Cancer Study (PanCan) lung cancer screening cohort (n = 2,357) and externally evaluated on the independent Evaluation of COPD Longitudinally to Identify Predictive Surrogate Endpoints (ECLIPSE) cohort (n = 1,753). AnatoMIL achieved competitive performance relative to strong MIL baselines, including a more complex transformer-based approach, while providing anatomically structured regional representations. Ablation experiments demonstrated complementary contributions from the anatomical hierarchy and ordinal severity objective, particularly for external severity assessment. The learned lobe contribution profiles revealed stage-dependent regional patterns that were consistent across both cohorts. These findings suggest that anatomy-aware MIL can provide a structured and interpretable representation of regional COPD involvement while maintaining predictive performance across cohorts.
Degree
thesis:*- Name thesis:degree_name
- Master of Applied Science - MASc
- Level thesis:degree_level
- master's
- Discipline thesis:degree_discipline
- Biomedical Engineering
- Grantor dc:publisher
- University of British Columbia
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Ahmadi, Raha
- Contributors dc:contributor
-
- Tam, Roger
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/2429/95309