University of Saskatchewan
Machine and Deep Learning for the Detection and Diagnosis of Ankylosing Spondylitis
Abstract
dc:description.abstractInformation is everywhere, especially in medical imaging. Radiomics refers to the generation of features from medical imagery, e.g., a patient’s scan frames, and its subsequent analysis via data-driven methods like machine and deep learning (ML/DL). A common radiomic application within medical imagery is disease prediction. Ankylosing spondylitis (AS) is an arthritis that manifests in the sacroiliac (SI) joints (in the pelvic region) and along the vertebral column. There is no cure, but when caught sufficiently early various treatments can be applied to lessen the disease’s impact on the patient’s daily activities. If left uncontrolled, the disease will lead to bone fusion (ankylosis), a painful condition which requires expensive surgeries. Unfortunately, AS has a lengthy average diagnostic delay of 5-10 years, resulting from various factors: the earliest symptom is lower back pain, which is common to many ailments; the earliest radiographic changes (inflammatory) are non-specific; the earliest bone changes (erosions) are very subtle; and AS is not well known outside musculoskeletal radiology. The estimated annual cost on the Canadian healthcare system is $13,000 CAD per AS patient, with costs increasing as the illness progresses to increasing severity. Herein this thesis, several radiomic models are developed, trained, and subsequently proposed, across numerous imaging modalities/sequences: CT, conventional MR (e.g., T1w, T2w), and UTE (a relatively recent type of MR that produces “pseudo-CT” imagery). In terms of diagnostic radiology, CT is best to image erosions, but cannot image inflammation. MR can image inflammation, but has diagnostic issues resulting in lower specificity. CT also requires radiation, which is unhealthy for patients, whereas MR does not. For CT, the best trained YOLOv5 SI joint localisation system achieved a 6-fold cross validation test set F1-score of 99.11%, mAP@0.5 of 99.27%, and mAP@0.5:0.95 of 65.10%. Using a randomised-shuffled dataset split, which simulates a wealth of data such that models were evaluated on similar but non-identical data to the training set, the best SI joint classification model was based on InceptionV3 and achieved an 8-fold cross-validation test set accuracy of 99.0%, sensitivity of 94.94%, specificity of 100.0%, and ROC AUC of 99.8%. Using a per-patient dataset split, which better exemplifies how models could perform in a real clinical setting given the training data, a hybrid Vision Transformer + InceptionV3 architecture was proposed for SI joint classification that achieved a 6-fold cross-validation test set accuracy of 87.04%, sensitivity of 69.54%, specificity of 96.02%, and ROC AUC of 90.6%. The results of the SI joint classifier were used to develop features to train and evaluate a series of traditional machine learning models for patient-level classification. The best performing patient-level models, Gaussian process or logistic regression, achieved a 6-fold cross-validation test set accuracy of 92.00%, AS sensitivity of 85.71%, erosion (early-stage AS) sensitivity of 75.00%, specificity of 95.38%, and ROC AUC of 86.90%. Over a reduced dataset wherein all patients had > 8 (the 25th percentile) of SI joints – to lessen the impact of SI joint misclassification errors – the best performing patient-level models achieved a 6-fold cross-validation test set accuracy of 98.63%, AS sensitivity of 96.43%, erosion sensitivity of 100.00%, specificity of 100.00%, and ROC AUC of 96.51%. Over this dataset, the system empowered a radiologist resident to improve their performance by correctly diagnosing 1 more AS and 1 more control patient. For MR, statistically significantly (p < 0.05) better results were achieved when combining different MR sequences together. The best trained YOLOv5 SI joint localisation system achieved a 6-fold cross-validation test set F1-score upwards 97.47%, mAP@0.5 upwards 97.52%, and mAP@0.5:0.95 upwards 54.30%. Using a randomised-shuffled dataset split, the best SI joint classification model was based on the first 9 blocks of ResNet50 and achieved a 6-fold cross-validation test set accuracy of 95.26%, sensitivity of 96.25%, specificity of 94.39%, and ROC AUC of 99.13%. Using a per-patient dataset split, the best SI joint classification system was comprised of a 2×VGG16-based model ensemble that achieved a 6-fold cross-validation test set accuracy of 86.84%, sensitivity of 83.33%, specificity of 95.00%, and ROC AUC of 89.17%. Human radiologists used the system’s predictions as a second reviewer to increase the specificities their diagnoses: 10% and 20% for an expert musculoskeletal and an expert general radiologist, respectively. For UTE, because a relatively small dataset was used, a small-scale pilot study was conducted. The best randomised-shuffled SI joint classification system that combined UTE with conventional MR was based on the first 6 blocks of InceptionV3 and achieved a 6-fold cross-validation test set accuracy of 98.84%, sensitivity of 98.39%, specificity of 99.09%, and ROC AUC of 99.9%. The best UTE-only randomised-shuffled SI joint classification system was based on the first 7 blocks of InceptionV3 and achieved a 6-fold cross-validation test set accuracy of 95.93%, sensitivity of 93.55%, specificity of 97.27%, and ROC AUC of 99.2%. UTE-trained model architectures generally outperform models trained on other single MR sequences. Given the current testing of the proposed AS classification systems, the patient-level MR system is recommended as the potentially most clinically useful system, whereas the patient-level CT pipeline is currently the most ready to undergo further clinical testing/evaluation and usage. Nevertheless, UTE – potentially combined with other conventional MR scan sequences – is believed to be the most promising research direction for automated AS diagnostics if and when a large(r) dataset of UTE scans is acquired and curated.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy (Ph.D.)
- Level thesis:degree_level
- Doctoral
- Discipline thesis:degree_discipline
- Electrical Engineering
- Department thesis:degree_department
- Electrical and Computer Engineering
- Grantor
- University of Saskatchewan
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Castro-Zunti, Riel D.
- Advisor dc:contributor.advisor
-
- Ko, Seok-Bum
- Committee members dc:contributor.committeemember
-
- Maleki, Farhad
- Stavness, Ian
- Chen, Li
- Teng, Daniel
- Wu, FangXiang
- Mohamed, Ebrahim
- ORCID person:identifier.orcid
Subjects
dc:subject × 6Rights
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/10388/17527