Date on Master's Thesis/Doctoral Dissertation
12-2025
Document Type
Doctoral Dissertation
Degree Name
Ph. D.
Department
Computer Engineering and Computer Science
Degree Program
Computer Science and Engineering, PhD
Committee Chair
Elmaghraby, Adel
Committee Co-Chair (if applicable)
El-Baz, Ayman
Committee Member
Zhang, Harry
Committee Member
Lauf, Adrian
Committee Member
Gentili, Monica
Author's Keywords
artifical intelligence; cancer imaging; carcinoma; pulmonary; nodules; hepatocellular
Abstract
Reliable, reproducible, and clinically useful artificial intelligence (AI) for oncologic imaging requires more than high headline metrics: it demands principled pipelines, calibrated predictions, and designs that generalize across anatomy, scanners, and institutions. This dissertation develops such pipelines in two complementary domains— hepatocellular carcinoma (HCC) on multiphase CT/MRI and indeterminate pulmonary nodules on thin-slice CT—while articulating evaluation and reporting practices that support credible clinical translation. The work proceeds in six core chapters. Two focused reviews synthesize the state of the art in (i) radiomics/AI for HCC and (ii) early assessment of lung nodules, mapping common pitfalls (data leakage, class imbalance, kernel dependence) and opportunities (calibration, selective prediction, external validation). Three technical chapters then instantiate and study end-to-end systems: (iii) a complete HCC computer-aided diagnosis pipeline that localizes liver tumors and grades them by fusing handcrafted radiomics with learned embeddings and calibrated decision layers; (iv) an integrated two-stage lungnodule segmentation framework that combines ROI localization with a 3D U-Net trained using an adaptive focal cross-entropy loss to sharpen boundaries under severe voxel imbalance; (v) a diagnosis model that leverages higher-order texture via a modified local ternary pattern alongside a stacking classifier to improve benign–malignant discrimination; and (vi) a brief thesis synopsis also outlines a transformer-fusion approach for lungnodule diagnosis that integrates HU, higher-order texture, and morphology descriptors while withholding manuscript-level detail to preserve journal submission. Across chapters, the thesis emphasizes: transparent preprocessing and fixed data splits; multiple, complementary validation regimes (e.g., Leave-One-Out and stratified k-fold); metrics beyond single numbers (ROC/PR with confidence intervals, Dice plus surface-distance measures); and calibration-first analysis (reliability diagrams, Brier score, expected calibration error). Empirically, better boundary fidelity in segmentation stabilizes downstream features and improves risk discrimination; stacking-based diagnosis consistently outperforms single models across validation settings; and calibrated probabilities enable clinically meaningful operating points and selective-prediction workflows. Error analyses highlight small-nodule and phase-timing challenges, as well as reconstruction-kernel sensitivity, motivating harmonization and shift-aware monitoring. The dissertation contributes unified designs that compose radiomics, deep learning, and higher-order texture into reliable pipelines; cross-organ principles for generalization and calibration; and practical guidelines for reporting, reproducibility, and workflow integration. Limitations of retrospective data and site heterogeneity are acknowledged, and a path is outlined toward prospective, multi-center validation and regulatory-grade evidence.
Recommended Citation
Alksas, Ahmed Muhammed Sherif Mostafa, "Artificial intelligence for cancer imaging: Integrated frameworks for hepatocellular carcinoma and pulmonary nodules." (2025). Electronic Theses and Dissertations. Paper 4682.
Retrieved from https://ir.library.louisville.edu/etd/4682
