Date of Award
2025-12-01
Degree Name
Doctor of Philosophy
Department
Electrical and Computer Engineering
Advisor(s)
Sergio D. Cabrera
Second Advisor
Zainul Abedin
Abstract
ABSTRACT
Study-I: An Image Processing Pipeline for Reference Guided Lung Region Detection in Chest Radiographs with Shape Similarity Matching Accurate and reliable segmentation of the lung region in chest X-ray (CXR) images is essential for computer-aided diagnosis (CAD) systems, particularly in the early detection and monitoring of lung disorders. Traditional segmentation techniques often rely on manual annotations, limiting scalability and adaptability. The proposed reference-guided approach selects the most similar healthy CXRs dynamically, ensuring flexibility while benefiting from dataset-specific reference images. Building upon prior approaches that utilize shape similarity-based selection and SIFT-flow, this study introduces a fully automated segmentation pipeline that eliminates the need for manually segmented lungs. Our method consists of three stages. First, for a given CXR, the five most similar CXR images are selected from a set of healthy CXR images using a partial Radon transform and Bhattacharyya shape similarity measures. Second, these identified healthy CXR images undergo segmentation through (a) Region of Interest (ROI) detection, (b) initial segmentation by performing gray-level Otsu thresholding using histogram of Canny edges within the ROI, (c) refinement of initial segmentation by merging the ROI gradient magnitude with the thresholded image, (d) salient point detection using the Gateaux derivative of total variation, and boundary adjustments via Bezier curve interpolation (e) morphological filtering for shape correction, and (f) final mask generation using the Graph Cut algorithm. Third, SIFT-flow first aligns each of the five segmented healthy CXR masks to the patient CXR, and the resulting patient-specific masks are then averaged to generate the final lung mask. Our approach ensures robust segmentation while minimizing reliance on expert annotations. Experimental evaluation across multiple datasets, including both normal and pathological cases, demonstrates high accuracy, sensitivity, and generalizability, making it a promising tool for automated lung segmentation in CAD applications.
Study-II: Patient-Specific Changes in T-wave Morphology for Acute Hyperkalemia Detection: A Paired-ECG Machine Learning Study Acute hyperkalemia is a common and potentially life threatening electrolyte disturbance in hospitalized patients, yet diagnosis still relies primarily on invasive blood sampling, which may be delayed and is vulnerable to pre analytical error, including hemolysis and sample handling artefacts. Characteristic T wave changes on the ECG are described, but conventional visual interpretation has limited sensitivity and is especially unreliable when potassium rises rapidly from a previously normal level. There is therefore a clear need for robust, non-invasive tools that can detect acute hyperkalemia. In this project we developed and evaluated a patient anchored, 12 lead ECG based method to detect acute hyperkalemia using quantitative T wave morphology. We retrospectively linked laboratory serum potassium measurements with diagnostic 12-lead ECGs from the MIMIC IV waveform and hospital databases, restricting to ECG–lab pairs obtained within one hour to capture acute electrolyte states. After excluding patients with chronic kidney disease, implanted pacemakers, significant arrhythmia, and hemolyzed samples, we identified adults who had both normokalemic ECGs, with serum potassium at or below 4.5 mmol/L, and acutely hyperkalemic ECGs, with serum potassium 5.7 mmol/L or higher, recorded within one hour of the blood test. Each ECG record was then passed through, custom signal processing and quality control pipeline, comprising baseline wander removal, noise filtering, R peak detection and automated localization of T wave onset, peak and offset in each lead. Applying these quality criteria and retaining only patients with at least one normokalemic and one hyperkalemic ECG record yielded a final cohort comprised of 500 patients contributing 1371 normokalemic and 558 hyperkalemic ECGs. Within each patient we then constructed “baseline anchored” delta features, such as changes in T-onset to T-peak duration, T wave duration, T wave full width at half maximum, T wave amplitude, T wave upstroke and R to T amplitude ratio, between a normokalemic reference ECG and candidate acutely hyperkalemic ECGs. These features were used to train logistic regression, support vector machine, random forest and XGBoost classifiers under group-wise cross validation with patient-level sample weighting and hyperparameter optimization. Across models, baseline anchored T wave deltas carried substantial discriminative information for acute hyperkalemia. XGBoost achieved the best out-of-fold performance, with a patient weighted area under the receiver operating characteristic curve (ROC AUC) of 0.94 and area under the precision–recall curve (PR AUC) of 0.92. At the threshold that maximized Youden’s J statistic, sensitivity was 89.6% and specificity 74.1%, corresponding to 500 true positive, 842 true negative, 58 false negative and 295 false positive predictions. The single feature analyses suggested that increases in T wave amplitude and upstroke, together with a fall in the R over T amplitude ratio, were the most informative markers. These findings demonstrate that a morphologically informed, baseline anchored analysis of T wave dynamics, combined with gradient boosted decision trees, can enable accurate ECG-based detection of acute hyperkalemia. Because the approach is patient specific ECG morphology rather than threshold based, it is well suited for real time deployment in monitors or ECG devices to provide continuous, non invasive surveillance for dangerous potassium elevations in high risk patients. Prospective validation and assessment of clinical impact will be important next steps before deployment into emergency and critical care practice.
Language
en
Provenance
Received from ProQuest
Copyright Date
2025-12
File Size
179 p.
File Format
application/pdf
Rights Holder
Basavarajaiah Shanmukhayya Totada
Recommended Citation
Shanmukhayya Totada, Basavarajaiah, "Image And Signal Processing Methods For Clinical Decision Support: Reference-Based Lung Segmentation On Chest X-Rays And Ecg-Based Study Of Acute Hyperkalemia Using Machine Learning" (2025). Open Access Theses & Dissertations. 4590.
https://scholarworks.utep.edu/open_etd/4590