Graduation Year

2026

Document Type

Dissertation

Degree

Ph.D.

Degree Name

Doctor of Philosophy (Ph.D.)

Degree Granting Department

Industrial and Management Systems Engineering

Major Professor

Jose Zayas-Castro, Ph.D.

Committee Member

Jorge Acuna, Ph.D.

Committee Member

Hadi Gard, Ph.D.

Committee Member

Ismail Uysal, Ph.D.

Committee Member

Jay Wolfson, Ph.D.

Keywords

Adaptive Pipeline Selection, Artificial Intelligence Robustness, Breast Ultrasound, CNN–Transformer Architectures, Healthcare Optimization, Medical Image Denoising

Abstract

Cancer care depends on timely and reliable decisions, from detection and diagnosis to treatment planning and patient monitoring. These decisions are often made under uncertainty because medical images and healthcare data may be noisy, incomplete, or difficult to interpret. In breast cancer imaging, ultrasound is widely used because it is safe, accessible, and complementary to other imaging modalities. However, variations in image quality, acquisition conditions, and noise can obscure lesion boundaries and texture, affecting human interpretation and artificial intelligence reliability. This dissertation develops deep learning, image-analysis, and optimization methods to improve healthcare decisions under imperfect information. Its primary focus is reliable breast ultrasound artificial intelligence through four connected studies of robustness evaluation, image restoration, adaptive pipeline recommendation, and interpretable perturbation-response analysis. Two additional studies extend this decision-support theme to hospital foodservice planning under competing objectives and uncertain patient demand.

The first study presents a noise-aware framework for evaluating a custom convolutional neural network and a pretrained Inception V3 model under controlled Gaussian, Poisson, and speckle noise across multiple severity levels. Although Inception V3 achieved stronger clean-image performance, both models experienced substantial degradation under severe noise. For example, the custom CNN’s accuracy decreased from 0.70 to 0.46 under severe Gaussian noise. Noise-matched training improved performance in several degraded conditions, but high-noise settings still reduced malignant recall and increased false-negative behavior. These findings demonstrate that clean-image accuracy alone does not adequately characterize clinical reliability and that robustness assessments should include class-specific outcomes.

The second study proposes DTRU-Net, a lightweight hybrid CNN–Transformer architecture that combines convolutional local-feature extraction, Transformer-based global-context modeling, and convolutional image reconstruction. The model was evaluated across the BUSI, BUS-UCLM, and MT Small datasets under Gaussian, Poisson, and speckle noise. DTRU-Net achieved its strongest gains under severe Gaussian and Poisson degradation, including an increase in BUS-UCLM classification accuracy from 0.33 to 0.57 under severe Poisson noise. More broadly, the model produced the best or tied-best classification performance in 11 of 12 evaluated settings. These results indicate that hybrid local–global modeling can improve image restoration while preserving information relevant to breast lesion classification.

The third study develops a multi-label recommendation framework that estimates the competence of four candidate diagnostic pipelines for each image-condition instance. On the held-out test set, the unweighted recommendation model achieved 94.02% practical diagnostic success, compared with 84.55% for the strongest fixed pipeline. Its performance also approached the 96.67% candidate-pool upper bound, showing that the candidate pipelines made complementary errors that could be exploited through adaptive selection. A more selective weighted model reduced malignant-case success, demonstrating that fewer recommendations do not necessarily provide safer diagnostic support.

The fourth study introduces a controlled speckle-response framework for identifying interpretable differences between benign and malignant lesion texture. Rather than treating speckle only as degradation, the study used controlled perturbation as a texture stress test. Malignant lesions consistently showed larger reductions in homogeneity and larger increases in contrast and dissimilarity than benign lesions. All nine feature-by-noise-level comparisons were statistically significant in BUSI, and the same directional patterns were reproduced in BUS-UCLM. These findings suggest that perturbation response can provide interpretable and reproducible information about lesion behavior that complements deep image representations.

The fifth study applies multi-objective optimization to hospital foodservice planning. The proposed model simultaneously reduced food waste and meal-building costs while maintaining patient nutritional requirements. In the case study, the multi-objective model reduced food waste by 19.70% and meal-building costs by 32.66% relative to the baseline. This demonstrates that environmental and financial objectives do not need to be addressed separately and can instead be optimized within a unified patient-centered planning framework.

The sixth study extends the hospital-foodservice research through a chance-constrained two-stage stochastic programming model that incorporates uncertain patient demand, last-minute purchases, surplus, waste, and nutritional reliability. In the case study, the proposed approach reduced food-storage costs by 60% and decreased annual food waste by approximately two tons. The analysis also revealed that the cost of increasing reliability rose sharply at the highest service levels, giving decision-makers a clearer understanding of the trade-off between operational protection and economic burden.

Overall, this dissertation contributes an integrated approach to healthcare decision support under imperfect conditions. The breast ultrasound studies demonstrate that medical artificial intelligence should be evaluated beyond clean-image accuracy and designed to remain informative under degradation, model variability, and lesion heterogeneity. The hospital foodservice studies show how uncertainty and competing nutritional, environmental, and financial objectives can be incorporated into operational decisions. Together, these contributions advance more robust, interpretable, and clinically relevant methods for transforming imperfect information into reliable healthcare decisions.

Download

Share

COinS