Carleton University
Techniques to Improve Classification Performance for Biomedical Applications with Limited Training Data and Variability
Abstract
dc:description.abstractReal-world data are often limited and can have heterogeneity due to sensor inconsistencies, where sensors manufactured by different entities, calibration, and sampling rates may be used to acquire the data. In the case of biomedical signals such as phonocardiograms or ultrasounds, signals may be acquired from multiple sites. Each site may present different signal characteristics. Further, it may not be possible to label signals cycle by cycle or images frame-wise, and only a consolidated record label for each acquisition may be available. Transferring record labels to cycles may lead to noisy labels. Learning to classify with limited data with variability due to heterogeneity and noisy labels is a hard problem, and this dissertation provides novel techniques for learning under such scenarios. Techniques for learning classifiers with limited data with variability in the presence and absence of a priori information are developed for 1D (phonocardiogram signals), 2D (breast ultrasound images) and 3D (lung ultrasound video clips) biomedical signals. To mitigate heterogeneity in the training data of 1D phonocardiogram signal, domain-balanced training (DBT), domain-balanced training with tracking (DBTT), domain-balanced training with cross-mix (DBCM), sequential decentralization-centralization aggregation strategy, and transformations such as Dynamic Mode Decomposition (DMD) are developed. These techniques utilize sensor information as part of the a priori information. For breast ultrasound 2D images, tumour masks are considered as a priori information and a novel domain knowledge training is proposed for a multiclass classifier. When no mask information is available in the training dataset, the mask information from those datasets that have masks is transferred through decentralized learning technique, through learnt weights to the binary breast cancer classifier. Frame-based selection utilizing signal processing metrics is proposed for mitigating data heterogeneity in 3D lung ultrasound video multiclass classification. Three methods, namely positive unlabelled training, average transformation and cycle-patching, are proposed to address variability due to noisy labels and train binary classifiers for 1D phonocardiogram signals. Significant improvements in classification results are observed when the various proposed training methods are compared with reimplemented state-of-the-art methods.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy (Ph.D.)
- Level thesis:degree_level
- Doctoral
- Discipline thesis:degree_discipline
- Engineering, Electrical and Computer
- Grantor dc:publisher
- Carleton University
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Ali, Ebrahim Ameen Ahmed
Subjects
dc:subject × 3Rights
dc:rights- Statement dc:rights
-
- Copyright © 2025 the author(s). Theses may be used for non-commercial research, educational, or related academic purposes only. Such uses include personal study, distribution to students, research and scholarship. Theses may only be shared by linking to the Carleton University Institutional Repository and no part may be copied without proper attribution to the author; no part may be used for commercial purposes directly or indirectly via a for-profit platform; no adaptation or derivative works are permitted without consent from the copyright owner.
- Language dc:language.iso
- en