Multi-Class Volumetric EM Segmentation using a 2D Model and Consistency-Enhancing Post-Processing

Abstract

Electron Microscopy (EM) data plays an important role in the reconstruction of neuronal and cellular structures through their segmentations. The segmentation of such structures provides great assistance in diagnosing diseases such as Alzheimer's and Parkinson's, and provides insight into other cellular structures, as well as white matter ultrastructures after brain injuries. However, the manual segmentation of EM data is time-consuming and task-intensive, since a thousand terabytes of images can be taken from a single cubic meter of tissue. Deep Learning (DL) has demonstrated strong performance in EM segmentation tasks across various studies; however, most of those studies focus on 2D single-class segmentation. This means that in real-life applications, if multiple structures need to be segmented, each structure would require a separate model, resulting in longer training time, more ground truth masks, and higher memory requirements to hold those models. Moreover, since the models are trained on single slices, they do not learn inter-slice features, which can result in inconsistent 3D-reconstructed structures with gaps. The latter can be addressed with 3D-based model training, but at the cost of higher computational requirements and heavier models. To address both limitations, a single 2D-based model can be trained to simultaneously segment all needed structures, followed by lightweight post-processing to enforce consistency in the segmented volume. There is limited exploration of such a pipeline in recent works, which is the aim of this thesis. In this thesis, we employ DeepLabV3+, a semantic segmentation model that is underexplored in the field of EM segmentation, for the task of dense segmentation of EM data, utilizing its ASPP module for multi-scale feature extraction. A benchmark of several segmentation architectures is conducted, including baseline segmentation models such as FCN and UNet, and more advanced models such as DeepLabV3+ and a UNet-like variant of it. Three lightweight post-processing filters are experimented with to enhance inter-slice segmentation consistency without the high-computational cost and overhead of 3D models. The results of this work show the effectiveness of this pipeline for the multi-class segmentation of cellular structures and provides a foundation for more future work on lightweight volumetric post-processing methods.

Date of Award	2026
Original language	American English
Awarding Institution	HBKU College of Science and Engineering