Graph Proposal Neural Networks for Cell Detection Immunostained Hodgkin Lymphoma Histology Images

The popularity of graph convolution networks has increased significantly in recent years due to their generality on non-Euclidean data. One of the areas in this realm that needs investigation is object detection. Due to the limits of clinical procedures, such as noisy imaging and ambiguity in the data, pathologists have difficulties in the early diagnosis of illness. In this instance, the automated clinical work of the medical examiners is made possible by the identification of various things in the cell image. Conventional convolutional neural networks (CNN) include various ways for processing Eu-clidean data for object detection tasks, but non-Euclidean data needs special consideration in this respect. In this research, we offer a region proposal technique for non-Euclidean data object recognition based on graph convolution. The extraction of a subgraph as a candidate for the potential object region, the categorization of object candidates into associated class labels, and the detection technique are our key areas of interest. For the immunostained images of lymphoma cancer patients’ blood cell dataset, we tested our technique against the euclidean domain Region Based Convolutional Neural Networks (R-CNN) method and discovered enhanced average precision and recall score by 12.8% and 34.72%.