A Framework for the Systematic Comparison of Classical Non-Learning, Machine Learning, and Artificial Intelligence Approaches to Cell Segmentation & Nuclei Detection in Histopathological Images

Abstract

In the fields of diagnosis, digital pathology, and drug discovery, the characterization of tissue, in particular the cells, is crucial because it provides a fundamental understanding of the underlying cellular dynamics and morphological information. For this reason, detecting cells from an image has become a pivotal step in most biomedical research prior to further analysis of relevant biological features, and it consists of complex and parameterized data transformation. The cell segmentation technique involves localising the cell object boundaries in a microscopic image domain and distinguishing these object boundaries from the background of the image. Objective and Rationale: Many different types of segmentation algorithms, classical, deep learning, etc.—may perform differently for different tissue types or different imaging modalities. Moreover, as different segmentation techniques have varying degrees of performance efficiency and accuracy for each benchmark dataset, choosing an appropriate segmentation algorithm remains a complicated matter. The purpose of this study is to establish a framework for the comparison and characterization of the performance of algorithms for the segmentation of cells and nuclei in histopathological images. This makes it possible to characterise under which circumstances one algorithm is performing better or worse than another, whether two algorithms are actually behaving differently or if the observed differences are really just artefacts of the testing set. To illustrate the usefulness of our framework, we will compare classical, non-learning algorithms against machine learning algorithms and a deep neural network (DNN) algorithm. Subjects and Methods: Publicly accessible and annotated data sets of H&E stained tissue images, along with widely agreed upon metrics to compare techniques such as the Jaccard index (JI), and Dice score coefficient (DSC) have been taken into consideration. A set of ground truth data consisting of manually outlined cell objects was used to compare with the cell segmentation masks for determining the outcome. For generalisation purposes, a different set of images was employed as the testing set to evaluate the performance and accuracy of the algorithm. Results and Discussion: The framework was effective in selecting the best-suited algorithm from the selected algorithms by using an adequate comparison of the performances of different algorithms and appropriate statistical analysis. The findings demonstrated that Random Forest, a machine learning-based algorithm, outperformed the other algorithms for the chosen dataset by yielding the highest score for all performance metrics in both the cases where the algorithms were compared with standard parameter values (DSC = 0.66) and when the algorithms were compared in various parameter spaces (DSC = 0.68). Statistical tests such as repeated measure ANOVA verified that the performances of each algorithm were statistically significantly distinct from one another and that there was a noticeable variance (73%) across the performance of different algorithms. The Shapiro-Wilk test was used to verify the normality of the data distribution, and post hoc tests were used to verify the repeated measure ANOVA's assumptions. With the use of several visualisation approaches, the dataset was characterised, revealing aspects that either enhanced or dropped an algorithm's effectiveness.