BE 645: Artificial Intelligence (AI) and Radiomics - Lecture 10 Recap (AI-Generated Podcast)

BE 645: Artificial Intelligence (AI) and Radiomics (Spring 2025)

The tenth lecture focuses on integrating artificial intelligence (AI) within a radiomics workflow, which encompasses sequential steps from image acquisition and region of interest (ROI) segmentation to radiomic feature extraction, preprocessing (selection/reduction), AI model development for diagnosis prediction, and rigorous model training, validation, and testing with performance metric reporting, culminating in explainability. A crucial aspect highlighted is data balancing to address imbalanced datasets where one class dominates, potentially biasing models. Techniques for data balancing include oversampling the minority class using methods like Random Oversampling (duplication, potentially leading to overfitting), Synthetic Minority Over-sampling Technique (SMOTE) (generating synthetic instances through interpolation), and Adaptive Synthetic (ADASYN) (focusing on underrepresented minority regions). Conversely, undersampling the majority class can be achieved via Random Undersampling (potential information loss), Cluster Centroids (representing clusters with centroids), and NearMiss (selecting near minority instances). Combining oversampling and undersampling (e.g., SMOTE + Tomek links or SMOTEENN) can sometimes be beneficial. Furthermore, the lecture emphasizes hyperparameter optimization, the critical process of selecting optimal hyperparameters (not learned from data) such as learning rate or number of layers. Various techniques exist, including Grid Search (exhaustive but costly), Random Search (more efficient sampling), Bayesian Optimization (probabilistic modeling), Gradient-based Optimization (for differentiable models), and Evolutionary Algorithms (e.g., genetic algorithms). Optuna, a powerful framework utilizing adaptive sampling and Bayesian optimization, is introduced for efficient hyperparameter tuning. The lecture outlines a workflow for building and tuning ML systems with hand-crafted features using Optuna, involving dataset handling, hyperparameter space definition, objective function, sampling, normalization, feature selection, model training, and result reporting. Finally, the lecture presents performance results (e.g., Weighted Accuracy, F1) achieved using Optuna to optimize models with different radiomic feature sets (First Order, GLCM, GLRLM, Shape, and All Features) in conjunction with various oversampling techniques, feature selection methods, and machine learning models, and provides a link to access the associated scripts.

Codes can be accessed using: https://github.com/HossamBalaha/BE-645-Artificial-Intelligence-and-Radiomics

Playlist from Spring 2025: https://www.youtube.com/playlist?list=PLVrN2LRb7eT0VBZqrtSAJQd2mqVtIDJKx

Old Playlist from Summer 2024: https://www.youtube.com/playlist?list=PLVrN2LRb7eT2KV3YMdXeF2B9dgaN4QF4g

Disclaimer: This is an AI-generated audio created using NotebookLM. The content may not reflect the opinions or expertise of a human author and is provided for informational or illustrative purposes only.

Видео BE 645: Artificial Intelligence (AI) and Radiomics - Lecture 10 Recap (AI-Generated Podcast) канала Hossam Magdy Balaha