Student name: Daniel S. Nicolau

RUN-EU institution: Politécnico de Leiria, Portugal

Abstract

Electron microscopy (EM) enables high-resolution imaging of cellular structures, such as mitochondria. However, the increasing volume of 3D EM datasets poses significant storage and transmission challenges. Lossy compression techniques can reduce data size; however, they risk degrading critical image details necessary for analysis. This study evaluates the impact of compression on machine vision models, specifically YOLOv8 (detection) and YOLOv8 + SAM (segmentation), when using H.265/HEVC and H.266/VVC standard codecs. Experimental results highlight that model performance degrades with higher compression rates. To mitigate this issue, a region-of-interest (ROI)-based coding strategy is proposed, preserving mitochondrial structures by applying different quantization parameters (QP) to foreground (ROIs) and background regions. Additionally, fine-tuning detection and segmentation models on compressed images improve their robustness to compression artefacts. The findings suggest that ROI-based coding, combined with model adaptation, can significantly enhance the efficiency of biomedical image storage and processing without compromising diagnostic accuracy, by almost reducing on average half of the data size for the same model’s performance.