Few-shot hypercolumn-based mitochondria segmentation in cardiac and outer hair cells in focused ion beam-scanning electron microscopy (FIB-SEM) data.

We present a novel AI-based approach to the few-shot automated segmentation of mitochondria in large-scale electron microscopy images. Our framework leverages convolutional features from a pre-trained deep multilayer convolutional neural network, such as VGG-16. We then train a binary gradient boosting classifier on the resulting high-dimensional feature hypercolumns. We extract VGG-16 features from the first four convolutional blocks and apply bilinear upsampling to resize the obtained maps to the input image size. This procedure yields a 2688-dimensional feature hypercolumn for each pixel in a 224 × 224 input image. We then apply L 1-regularized logistic regression for supervised active feature selection to reduce dependencies among the features, to reduce overfitting, as well as to speed-up gradient boosting-based training. During inference we block process 1728 × 2022 large microscopy images. Our experiments show that in such a formulation of transfer learning our processing pipeline is able to achieve high-accuracy results on very challenging datasets containing a large number of irregularly shaped mitochondria in cardiac and outer hair cells. Our proposed few-shot training approach gives competitive performance with the state-of-the-art using far less training data.

Automatic segmentation of mitochondria in EM data using pairwise affinity factorization and graph-based contour searching.

In this paper, we investigate the segmentation of closed contours in subcellular data using a framework that primarily combines the pairwise affinity grouping principles with a graph partitioning contour searching approach. One salient problem that precluded the application of these methods to large scale segmentation problems is the onerous computational complexity required to generate comprehensive representations that include all pairwise relationships between all pixels in the input data. To compensate for this problem, a practical solution is to reduce the complexity of the input data by applying an over-segmentation technique prior to the application of the computationally demanding strands of the segmentation process. This approach opens the opportunity to build specific shape and intensity models that can be successfully employed to extract the salient structures in the input image which are further processed to identify the cycles in an undirected graph. The proposed framework has been applied to the segmentation of mitochondria membranes in electron microscopy data which are characterized by low contrast and low signal-to-noise ratio. The algorithm has been quantitatively evaluated using two datasets where the segmentation results have been compared with the corresponding manual annotations. The performance of the proposed algorithm has been measured using standard metrics, such as precision and recall, and the experimental results indicate a high level of segmentation accuracy.

Unsupervised mitochondria segmentation using recursive spectral clustering and adaptive similarity models.

The unsupervised segmentation method proposed in the current study follows the evolutional ability of human vision to extrapolate significant structures in an image. In this work we adopt the perceptual grouping strategy by selecting the spectral clustering framework, which is known to capture perceptual organization features, as well as by developing similarity models according to Gestaltic laws of visual segregation. Our proposed framework applies but is not limited to the detection of cells and organelles in microscopic images and attempts to provide an effective alternative to presently dominating manual segmentation and tissue classification practice. The main theoretical contribution of our work resides in the formulation of robust similarity models which automatically adapt to the statistical structure of the biological domain and return optimal performance in pixel classification tasks under the wide variety of distributional assumptions.