Egilea: Mohamad Abou Ali

Izenburua: Advancing artificial intelligence in medical imaging: self-aware exploration, data augmentation, and real- world simulation for robust diagnosis

Zuzendariak: Fadi Dornaika eta Alireza Bosaghzadeh

Eguna: 2025ko maiatzaren 2an
Ordua: 10:30h
Lekua: Zuzenbideko fakultateko gradu-aretoan

Abstract:

"Artifícial intelligence (Al) and deep learning are at the forefront of revolutionizing medica! imaging, promising unprecedented advancements in diagnostic precision, efficiency, and accessibility. Despite this potential, their application in clinical settings is hampered by persistent challenges such as limited high-quality data, class imbalance, and the complex variability inherent in medical images. This PhD thesis addresses these critica! challenges through three pioneering research contributions: (1) the introduction of self-aware Al systems capa ble of autonomous learning optimization, (2) the creation of a novel data augmentation technique that enhances model performance under constrained data conditions, and (3) the development of robust deep learning frameworks tailored far MRlbased brain cancer diagnosis in real-world scenarios.

The first key innovation is the exploration of self-aware Al. This thesis propases that pre-trained models such as Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) possess the potential to self-monitor and adapt their learning processes autonomously. By introducing novel metrics, such as "accuracy difference" and "loss difference," this research offers a framework for measuring and enhancing the self-awareness of Al models. This development has profound implications for the future of Al in healthcare, paving the way for adaptive models that can fine-tune themselves in response to the quality and quantity of the data they encounter, ensuring more reliable and context-aware decision-making in clinical environments.

The second contribution is the introduction of the 'Naturalize' augmentation technique, a groundbreaking method designed to generate synthetic medical images with fidelity equal to original data. This technique is specifically engineered to overcome data scarcity and class imbalance, two significant obstacles in medical imaging. By applying 'Naturalize' to blood cell and skin cancer classification, this work demonstrates unparalleled improvements in both sensitivity and specificity, particularly in underrepresented classes. The 'Naturalize' technique fundamentally transforms the dataset augmentation paradigm, enabling deep models to generalize more effectively to unseen clinical cases.

The third contribution focuses on advancing model robustness for MRl-based brain cancer diagnosis. Medical imaging, particularly MRI, is subject to significant variability dueto differences in magnetic field strength, patient motion, and image quality. This work systematically addresses these challenges by simulating realworld conditions-such as noise, blur, and motion artifacts-within the training process. By integrating sophisticated data augmentation strategies, including Gaussian noise and blur, into the training of deep learning models, the research significantly enhances model resilience and generalization, ensuring consistent and accurate performance across diverse clinical environments. This contribution is a critica! step toward translating Al solutions from research to practice in the doma in of neuro-oncology.

In summary, this thesis makes substantial strides in addressing sorne of the most pressing limitations of Al in medical imaging. Through innovative augmentation techniques, the concept of self-aware Al, and robust training methodologies, this work significantly enhances the reliability, adaptability, and clinical relevance of deep learning models. The findings lay the foundation for more intelligent, autonomous, and clinically applicable Al systems that can improve diagnostic accuracy and patient outcomes. Future research directions include expanding the 'Naturalize' method to 3D imaging, developing real-time adaptive Al models, and exploring federated learning to enhance model generalization while preserving patient privacy.

K.eywords: Deep Learning, Transfer Learning, Fine-tuning, Medical lmage, Augmentation, Naturalize, GANs, Pre-trained Models, Medical lmaging, Selfawareness, Model Robustness, Real-Life Scenario Simulation."