Survival analysis from medical images using machine learning

Survival analysis, which examines the time until a specific event occurs, is widely used in various fields such as engineering and social sciences. However, it is particularly crucial in disease prognosis, treatment efficacy evaluation, and identifying risk factors. This thesis provides a detailed examination of the fundamental concepts of survival analysis, a taxonomy of commonly used methods, and the tools utilized in the analysis. A systematic review of studies conducted between 2020 and 2024 on overall survival in lung cancer using artificial intelligence techniques has been carried out. To support researchers working on survival analysis, a specially designed GPT-based tool has been introduced. Additionally, a novel feature called GTV1-SliceNum, which considers the number of tumor-containing slices in a patient, has been proposed, and its contribution to survival classification performance has been demonstrated. Furthermore, a new loss function, PEN-BCE, which penalizes false negatives and false positives, has been developed and validated using publicly available cancer datasets to assess its impact on classification performance. Finally, survival analysis problems have been approached using different data types and methodologies, with comprehensive performance evaluations conducted through ablation studies. Experimental results indicate that the proposed GTV1-SliceNum feature and PEN-BCE loss function provide significant performance improvements compared to existing methods. This study offers an innovative contribution to survival analysis problems from both theoretical and practical perspectives and proposes potential approaches that can be integrated into medical decision support systems.