Leveraging Machine Learning to Unravel Autophagy Dynamics in Cellular Biology

Abstract

Autophagy, a vital biological process for maintaining cellular homeostasis, plays a crucial role in various health conditions. However, manual annotation and analysis of fluorescence microscopy images, commonly used in autophagy research, are labor-intensive and error-prone. Leveraging artificial intelligence presents a promising solution to automate and enhance these tasks, potentially offering new insights into autophagy-related phenomena. For this purpose, the thesis investigates the potential of artificial intelligence in analyzing specific cellular imaging data. It covers the development and training of models, as well as interpreting the results, which is crucial for uncovering potentially relevant biological insights.

This thesis conducts a comprehensive analysis of the CELLULAR data set using deep learning techniques. The data set consists of time-lapse images of cells undergoing autophagy. Multiple well-established deep learning models were trained and evaluated for the detection, segmentation, classification, and tracking of cells. Explainable artificial intelligence techniques were employed to enhance understanding of the data and uncover new insights into autophagy, while also ensuring the reliability of the model's results and their ethical contribution to research.

The findings indicate that combining object detection with segmentation produces superior results compared to segmentation alone. The fine-tuned vision transformer outperformed convolutional networks in cell classification, achieving an accuracy of 86\% in distinguishing between three types of cells. Similarly, U-Net++ achieved the same level of accuracy in cell segmentation. Generative models, particularly the diffusion model, showed promising results in predicting final frames within image sequences.

Furthermore, the top-performing models were integrated into a desktop application, facilitating image uploading and automated data annotation generation. These annotations mimic the format of the original CELLULAR data set, comprising bounding box coordinates, segmentation masks, and class labels.

Working closely with biology experts throughout this thesis refined the focus of the research and helped interpret findings. The results on model explainability revealed the aspects that the classification model emphasized during decision-making, which the biologists verified as relevant factors. This collaborative approach increased the relevance and practical application of the research findings.