Semantic segmentation of HeLa cells: An objective comparison between one traditional algorithm and four deep-learning architectures

Abstract

The quantitative study of cell morphology is of great importance as the structure and condition of cells and their structures can be related to conditions of health or disease. The first step towards that, is the accurate segmentation of cell structures. In this work, we compare five approaches, one traditional and four deep-learning, for the semantic segmentation of the nuclear envelope of cervical cancer cells commonly known as HeLa cells. Images of a HeLa cancer cell were semantically segmented with one traditional image-processing algorithm and four three deep learning architectures: VGG16, ResNet18, Inception-ResNet-v2, and U-Net. Three hundred slices, each 2000 × 2000 pixels, of a HeLa Cell were acquired with Serial Block Face Scanning Electron Microscopy. The first three deep learning architectures were pre-trained with ImageNet and then fine-tuned with transfer learning. The U-Net architecture was trained from scratch with 36, 000 training images and labels of size 128 × 128. The image-processing algorithm followed a pipeline of several traditional steps like edge detection, dilation and morphological operators. The algorithms were compared by measuring pixel-based segmentation accuracy and Jaccard index against a labelled ground truth. The results indicated a superior performance of the traditional algorithm (Accuracy = 99%, Jaccard = 93%) over the deep learning architectures:

Volumetric Semantic Instance Segmentation of the Plasma Membrane of HeLa Cells

“…For the cell not including the nucleus AC = 0.9629, J I = 0.8094, for the cell and the nucleus AC = 0.9655, J I = 0.8711 and for the nucleus alone AC = 0.9975, J I = 0.9665. The algorithm to segment the nucleus provided excellent results, and it had previously been reported that it outperformed several deep learning architectures [59]. The small differences between the segmented nucleus and that of a manual expert segmentation are due mainly to the calculations of the thickness of the NE and small invaginations (Figures 12, 13 right column).…”

“…This paper describes an extension to previous work which focused on the segmentation the NE of a cell from a cropped volume [59,60]. In this work individual HeLa cells and their nuclei are instance segmented in 3D.…”

“…A notable contribution is CEM500K [58], a very large dataset of images, pre-trained models and curation pipeline for model building specific for Electron Microscopy. In addition, the specific nature of a semantic segmentation can allow traditional algorithms to provide satisfactory results, in some cases superior to deep learning approaches [59].…”

CEM500K, a large-scale heterogeneous unlabeled cellular electron microscopy image dataset for deep learning

“…Transfer learning, when relevant pre-trained parameters are available, is the default approach for extracting the best performance out of small training datasets ( Huh et al, 2016 ; Devlin et al, 2018 ). While ImageNet pre-trained models are sometimes used for cellular EM segmentation tasks ( Karabağ et al, 2020 ; Devan et al, 2019 ), high-level features learned from ImageNet may not be applicable to biological imaging domains ( Raghu et al, 2019 ). Building a more domain-specific annotated dataset large enough for pre-training would be a significant bottleneck, and indeed, it required multiple years to annotate the 3.2 × 10 6 images that form the basis of ImageNet.…”

CEM500K – A large-scale heterogeneous unlabeled cellular electron microscopy image dataset for deep learning

“…Transfer learning, when relevant pre-trained parameters are available, is the default approach for extracting the best performance out of small training datasets [27] [28]. While ImageNet pre-trained models are sometimes used for cellular EM segmentation tasks [29] [30], high-level features learned from ImageNet may not be applicable to biological imaging domains [31].Building a more domain-specific annotated dataset large enough for pre-training would be a significant bottleneck, and indeed, it required multiple years to annotate the 3.2 x 10 6 images that form the basis of ImageNet. Fortunately, recent advances in unsupervised learning algorithms have now enabled effective pre-training and transfer learning without the need for any up front annotations; in fact, on many tested benchmarks, unsupervised pre-training leads to better transfer learning performance [32][33] [34] [35] [36][37] [38].…”