Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
CACO-phony: contrastive architecture for TB cough classification
Thesis (MEng)--Stellenbosch University, 2025.
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English |
| Published: |
Stellenbosch : Stellenbosch University
2025
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| _version_ | 1867613953185546240 |
|---|---|
| access_status_str | Open Access |
| author | Farrell, Minette |
| author2 | Niesler, T. R. (Thomas) |
| author_browse | Farrell, Minette Niesler, T. R. (Thomas) |
| author_facet | Niesler, T. R. (Thomas) Farrell, Minette |
| author_sort | Farrell, Minette |
| collection | Thesis |
| dc_rights_str_mv | Stellenbosch University |
| description | Thesis (MEng)--Stellenbosch University, 2025. |
| format | Thesis |
| id | oai:scholar.sun.ac.za:10019.1/134616 |
| institution | Stellenbosch University (South Africa) |
| language | English |
| last_indexed | 2026-06-10T12:44:19.493Z |
| license_str | Other — see source repository |
| provenance_str_mv | Harvested via OAI-PMH from SUNScholar — Stellenbosch University Repository |
| publishDate | 2025 |
| publishDateRange | 2025 |
| publishDateSort | 2025 |
| publisher | Stellenbosch : Stellenbosch University |
| publisherStr | Stellenbosch : Stellenbosch University |
| record_format | dspace |
| source_str | SUNScholar — Stellenbosch University Repository |
| spelling | oai:scholar.sun.ac.za:10019.1/134616 CACO-phony: contrastive architecture for TB cough classification Farrell, Minette Niesler, T. R. (Thomas) Stellenbosch University. Faculty of Engineering. Dept. of Electrical & Electronic Engineering. Tuberculosis -- Diagnosis Computer-aided diagnosis Artificial intelligence -- Medical applications Respiratory organs -- Sounds Thesis (MEng)--Stellenbosch University, 2025. Farrell, M. 2025. CACO-phony: Contrastive Architecture for TB Cough classification. Unpublished masters thesis. Stellenbosch: Stellenbosch University [online]. Available: https://scholar.sun.ac.za/items/8cd4a3e0-7f23-4d47-ba89-789d449bad6f ENGLISH ABSTRACT: Is it possible to distinguish TB positive patients from TB negative patients, based solely on the sound of their coughs? Despite being both preventable and treatable, tuberculosis is one of the leading causes of death worldwide and still poses a major threat to society, especially in developing countries. While medical professionals assert that they cannot distinguish TB positive patients from TB negative patients by listening to their coughs, machine learning models are able to extract some aspect of the cough that allows it to perform TB classification. Contrastive learning is a machine learning technique that is used to learn robust and informative features from the data. This project aims to determine whether contrastive learning can be used to generate features from the cough audio data to better perform TB classification. The presented work uses a Siamese network with contrastive loss functions to perform audio classification using cough audio data from TB positive and negative patients. Supervised contrastive loss functions (pair loss and triplet loss) and self-supervised loss functions (NT-Xent loss and Barlow twins loss) are used to perform contrastive learning. In addition, a new supervised contrastive loss function is proposed, named exclusive loss. Contrastive architectures are implemented using AlexNet, VGGish and BiLSTM models as encoders. The results show that a contrastive architecture using VGGish offers the best performance among the three contrastive architectures, while the novel exclusive loss delivers the best performance among the five contrastive loss functions. VGGish trained non-contrastively with cross-entropy loss delivered a test AUC (mean and standard deviation) of 0.637± 0.071. The contrastive architecture with VGGish trained with the novel exclusive loss improved this baseline result with a test AUC (mean and standard deviation) of 0.653±0.093. When VGGish is further pre-trained contrastively with Barlow twins loss on the Coswara dataset, the test AUC of the contrastive architecture with VGGish improves to 0.665 ± 0.074. AFRIKAANSE OPSOMMING: Is dit moontlik om TB-positiewe pasi¨ente van TB-negatiewe pasi¨ente to onderskei gebasseer slegs op hulle hoes? Tuberkulose is een van die hoofoorsake van sterftes wˆereldwyd en dit is nogsteeds ’n bedreiging tot die samelewing, veral in ontwikkelende lande, ten spyte daarvan dat dit genees kan word. Dokters beweer dat hulle nie ’n pasi¨ent kan diagoniseer slegs gebasseer op hulle hoes nie. Masjienleermodelle kan sekere aspekte van ’n hoes identifiseer wat dit toelaat om onderskeid te tref tussen TB-positiewe en TB-negatiewe pasi¨ente. ’n Masjienleertegniek genaamd kontrasterende leer word dikwels gebruik om insiggewende kenmerke te leer vanuit die data. Die doel van hierdie projek is om te identifiseer of kontrasterende leer gebruik kan word om kenmerke te leer van die hoesopnames om TB te klassifiseer. Die model gebruik ’n Siamese netwerk met ’n kontrasterende leerfunksie om oudio klassifisikasie te doen met die hoes opnames van TB-positiewe en TB-negatiewe pasi¨ente. Toesighoudende kontrasterende verliesfunksies (pair loss en triplet loss) en selftoesighoudende kontrasterende verliesfunksies (NT-Xent loss en Barlow twins loss) word gebruik om die masjienleermodelle kontrasterend te leer. Verder word ’n nuwe toesighoudende kontrasterence verliesfunksie genoemd exclusive loss ook voorgestel. Die kontrasterende argitekture word ge¨ımplementeer deur AlexNet, VGGish en BiLSTM modelle te gebruik as enkodeerder modelle. Die resultate toon aan dat die kontrasterende argitektuur met VGGish die beste presteer uit die drie argitekture, en dat die nuwe exclusive loss die beste presteer uit die vyf kontrasterende verliesfunksies. VGGish sonder kontrasterendleer het op die toetsstel ’n AUC (gemiddeld met standaard afwyking) van 0.637 ± 0.071 verower. Die kontrasterende argitektuur met VGGish wat met die nuwe exclusive loss geleer is het ’n AUC (gemiddeld met standaard afwyking) van 0.653 ± 0.093 verower. VGGish wat verder geleer was met die kontrasterende Barlow twins loss op die Coswara datastel, het die kontrasterende argitektuur se AUC verbeter tot 0.665 ± 0.074. Masters 2025-12-18T13:33:20Z 2025-12-18T13:33:20Z 2025-12 Thesis https://scholar.sun.ac.za/handle/10019.1/134616 en Stellenbosch University xxiii, 139 pages : illustrations application/pdf Stellenbosch : Stellenbosch University |
| spellingShingle | Tuberculosis -- Diagnosis Computer-aided diagnosis Artificial intelligence -- Medical applications Respiratory organs -- Sounds Farrell, Minette CACO-phony: contrastive architecture for TB cough classification |
| title | CACO-phony: contrastive architecture for TB cough classification |
| title_full | CACO-phony: contrastive architecture for TB cough classification |
| title_fullStr | CACO-phony: contrastive architecture for TB cough classification |
| title_full_unstemmed | CACO-phony: contrastive architecture for TB cough classification |
| title_short | CACO-phony: contrastive architecture for TB cough classification |
| title_sort | caco phony contrastive architecture for tb cough classification |
| topic | Tuberculosis -- Diagnosis Computer-aided diagnosis Artificial intelligence -- Medical applications Respiratory organs -- Sounds |
| url | https://scholar.sun.ac.za/handle/10019.1/134616 |
| work_keys_str_mv | AT farrellminette cacophonycontrastivearchitecturefortbcoughclassification |