Background:
Structural remodeling is a hallmark of cardiovascular disease (CVD). Electron microscopy (EM) imaging is commonly used to characterize subcellular, ultrastructural remodeling in CVD. While technical advances have enabled the acquisition of larger datasets, analysis workflows still rely largely on subjective description and procedures such as manual annotation for quantitative analysis. This limits efficiency and reproducibility. Moreover, most current workflows do not distinguish individual subcellular structures, restricting objective analysis to proportional area measurements.
Methods:
Murine atrial 2D EM images (3400×) prepared using standard histological procedures were annotated for subcellular structures. Annotation was performed either manually using conventional software or interactively in a custom platform assisted by the pretrained Segment Anything Model 2 (SAM2). SAM2-assisted annotations were interactively refined by positive and negative point prompts. The performance of SAM2-assisted annotation was evaluated against manual annotations using intersection over union (IoU) and expert review. 200 annotations of the same sarcomere structures, mitochondria, and lipofuscin droplets were randomly selected. For expert review, three reviewers blinded to the annotation method independently rated paired annotations shown side by side. A custom nnU-Net model was trained on 35 annotated images. Using this model, a fully automated semantic segmentation of a previously published, manually analyzed dataset was performed and compared with the prior manual analysis.
Results:
SAM2-assisted annotation reduced annotation time by at least 50% compared with the reference manual workflow. Annotators reported high rates of sufficiently accurate annotation of comparatively simple structures such as mitochondria after a single point prompt. The mean IoU between manual and SAM2-assisted annotations was 90.4±3.1%, 68.2±13.4%, and 83.1±6.8% for mitochondria, sarcomeres, and lipofuscin droplets, respectively. In expert review, SAM2-assisted annotations were preferred in 82.6±5.0%, 72.9±7.4%, and 69.0±10.7% of cases for mitochondria, sarcomeres, and lipofuscin droplets, respectively. Interrater reliability was high (ICC for absolute agreement = 0.791, p<0.05). Evaluators particularly noted improved delineation of organelle boundaries and better separation of closely adjacent objects. Fully automated semantic segmentation reproduced previously reported group differences in proportional area measurements in the re-analyzed dataset.
Conclusion:
We present a machine learning-assisted annotation tool for quantitative analysis of image datasets. This tool improves both time efficiency and annotation quality compared with our reference manual workflow. It also facilitates the generation of training datasets for machine learning and supports the scalable analysis of growing datasets. Results from a previously analyzed dataset were reproduced in only a fraction of the original analysis time. This provides the basis for a future high-throughput pipeline for automated instance segmentation and classification, enabling objective and reproducible organelle-level analysis of subtle tissue remodeling.
Supported by the DFG and Deutsche Herzstiftung
