Development of a damage detection method in DNA images using convolutional neural networks

Human cells are continuously subjected to attacks from both internal (endogenous) and external (exogenous) factors. These factors can alter the chemical structure of nucleotides or disrupt the phosphodiester backbone of DNA, leading to DNA damage. In every cell, thousands of DNA damage events occur simultaneously, threatening genomic integrity. Accurate understanding and classification of DNA damage are crucial in fields such as genetics, cancer research, and toxicology. One widely used method for detecting DNA damage is the Comet Assay, also known as Single-Cell Gel Electrophoresis. This tech-nique allows for the visualization of DNA damage at the single-cell level. During electro-phoresis, damaged DNA migrates out of the nucleus, forming a tail resembling a comet, while intact DNA remains compact. The length and intensity of the tail are proportional to the amount of DNA damage, providing a sensitive and effective means for measuring DNA fragmentation. With advancements in technology, DNA damage analysis has in-creasingly incorporated artificial intelligence (AI) and machine learning (ML) techniques. These technologies not only improve the speed and accuracy of DNA damage detection and classification but also facilitate the identification of damage types and the prediction of repair outcomes. AI and ML models can be trained on large datasets to identify pat-terns specific to various types of DNA damage. These models utilize deep learning archi-tectures such as convolutional neural networks (CNNs) to analyze and interpret complex data. In this thesis, various methods and models are proposed for the classification of DNA damage observed in microscope images using deep learning-based (ESA) approach-es. The study includes the labeling of DNA images, the classification of these labeled images using current deep learning architectures from the literature, the evaluation of clas-sification results, the optimization of hyperparameters for models achieving the highest performance, and the retraining of all architectures based on optimized parameters, fol-lowed by a comparison of performance metrics. In the second phase, models were devel-oped to achieve the most accurate DNA damage classification through hyperparameter optimization with the Keras library. The performance of these models was compared, and a 20-layer CNN model achieved the highest accuracy of 89%. The developed model out-performed the currently used ESA models in terms of accuracy, and further optimization of ESA models resulted in even higher performance levels. Using the developed model, object recognition applications were implemented, enabling automatic classification of DNA damage images. Additionally, dataset augmentation was performed to further en-hance the performance of the developed model. This study aims to advance DNA dam-age classification with deep learning methods, ultimately improving the automation and accuracy of DNA damage analysis.