Abstract:
Etioplasts are plastids that develop in dark grown tissues and later convert into chloroplasts when exposed to light. Their internal ultrastructure made up of the prolamellar body (PLB), prothylakoids (PTs), plastoglobules and occasional starch grains contains important information about plastid development, stress responses and genetic changes. Transmission electron microscopy (TEM) is the gold standard for visualising these nanostructures, but quantitative analysis still depends largely on manual tracing in tools such as ImageJ/Fiji. This makes large scale studies slow, subjective and difficult to reproduce. This thesis presents an automated framework for quantitative analysis of etioplast ultrastructure from TEM micrographs. TEM images of Arabidopsis thaliana dark grown seedlings, acquired in collaboration with a CNRS laboratory, are manually annotated and then validated by a plastid expert to obtain high quality ground truth masks for etioplast envelopes, PLBs, PTs, plastoglobules and starch grains. After standardised preprocessing and data augmentation, these annotations are used to train YOLO based instance segmentation models (YOLOv5-seg, YOLOv8-seg and YOLOv12-seg). The best performing model is exported as a .pt checkpoint and integrated into a server side pipeline that accepts new TEM images via an API, runs inference and extracts morphometric descriptors such as etioplast and PLB area, PT number and total length, plastoglobule count and plastoglobule diameter. All measurements are stored in structured CSV or JSON files for downstream analysis. A lightweight generative AI module then operates only on these numeric outputs, producing short, human-readable summaries without altering any underlying measurements. The system is evaluated using standard segmentation metrics (Dice, IoU, precision, recall), parameter accuracy measures against expert annotations (MAE, RMSE, MAPE, Bland Altman analysis) and qualitative expert review. Overall, the results indicate that modern segmentation models, combined with explicit calibration and quality control, can turn TEM-based etioplast analysis from a manual, artisanal procedure into a reproducible and scalable quantitative pipeline
