PROJECT TITLE :

Bayesian Polytrees With Learned Deep Features for Multi-Class Cell Segmentation

ABSTRACT:

Quantitative cell biology relies heavily on being able to identify different cell compartments, cell types, and the ways in which they interact. Although this job can be automated, it is difficult because of the high resemblance of objects from other classes and irregularly shaped structures that make this task difficult. For example, graphical models are useful because they may employ prior knowledge and model inter-class dependencies to help relieve the problem The use of directed acyclic graphs, such as trees, to model top-down statistical dependences as a prior for enhanced picture segmentation has been widely used. Trees, on the other hand, can capture some inter-class limitations. While tree-based techniques can capture label proximity relations, polytree graphical models can do so more organically. An efficient method for calculating closed-form posteriors of network nodes on polytrees has been created using a two-pass message transmission mechanism. It is tested on simulated and public fluorescence microscopy datasets, outperforming three state-of-the-art convolutional neural networks, SegNet, DeepLab, and PSPNet, in terms of performance. By highlighting portions in the segmented image that do not conform to prior knowledge, polytrees beat directed trees in forecasting segmentation error. In this method, it is possible to quantify the degree of uncertainty in the segmentation results and use this information to improve the segmentation further.