MOTIVATION: ITree is an intuitive web tool for the manual, semi-automatic, and automatic induction of decision trees. It enables interactive modifications of tree structures and incorporates Relative Expression Analysis for detecting complex patterns in high-throughput molecular data. This makes ITree a versatile tool for both research and education in biomedical data analysis. RESULTS: The tool allows users to instantly see the effects of modifications on decision trees, with updates to predictions and statistics displayed in real time, facilitating a deeper understanding of data classification processes. AVAILABILITY AND IMPLEMENTATION: Available online at https://itree.wi.pb.edu.pl. Source code and documentation are hosted on GitHub at https://github.com/hsokolowski/iTree and in supplement.

Decision trees (DTs) are popular techniques in the field of eXplainable Artificial Intelligence. Despite their effectiveness in solving various classification problems, they are not compatible with modern biological data generated with high-throughput technologies. This work aims to combine evolutionary induced DT with a recently developed concept designed directly for gene expression data, called Relative eXpression Analysis (RXA). We propose a new solution, termed Evolutionary Heterogeneous Decision Tree (EvoHDTree), which uses both classical univariate and bivariate tests that focus on the relative ordering and weight relationships between the genes in the splitting nodes. The search for the decision tree structure, node representation, and splits is performed globally by the evolutionary algorithm.

Abstract This paper concerns the evolutionary induction of decision trees (DT) for large-scale data. Such a global approach is one of the alternatives to the top-down inducers. It searches for the tree structure and tests simultaneously and thus gives improvements in the prediction and size of resulting classifiers in many situations. However, it is the population-based and iterative approach that can be too computationally demanding to apply for big data mining directly. The paper demonstrates that this barrier can be overcome by smart distributed/parallel processing. Moreover, we ask the question whether the global approach can truly compete with the greedy systems for large-scale data. For this purpose, we propose a novel multi-GPU approach. It incorporates the knowledge of global DT induction and evolutionary algorithm parallelization together with efficient utilization of memory and computing GPU’s resources. The searches for the tree structure and tests are performed simultaneously on a CPU, while the fitness calculations are delegated to GPUs. Data-parallel decomposition strategy and CUDA framework are applied. Experimental validation is performed on both artificial and real-life datasets. In both cases, the obtained acceleration is very satisfactory. The solution is able to process even billions of instances in a few hours on a single workstation equipped with 4 GPUs. The impact of data characteristics (size and dimension) on convergence and speedup of the evolutionary search is also shown. When the number of GPUs grows, nearly linear scalability is observed what suggests that data size boundaries for evolutionary DT mining are fading.

Abstract In this paper, we present a fast multi-stage image segmentation method that incorporates texture analysis into a level set-based active contour framework. This approach allows integrating multiple feature extraction methods and is not tied to any specific texture descriptors. Prior knowledge of the image patterns is also not required. The method starts with an initial feature extraction and selection, then performs a fast level set-based evolution process and ends with a final refinement stage that integrates a region-based model. The presented implementation employs a set of features based on Grey Level Co-occurrence Matrices, Gabor filters and structure tensors. The high performance of feature extraction and contour evolution stages is achieved with GPU acceleration. The method is validated on synthetic and natural images and confronted with results of the most similar among the accessible algorithms.

Relative Expression Analysis (RXA) focuses on finding interactions among a small group of genes and studies the relative ordering of their expression rather than their raw values. Algorithms based on that idea play an important role in biomarker discovery and gene expression data classification. We propose a new evolutionary approach and a paradigm shift for RXA applications in data mining as we redefine the inter-gene relations using the concept of a cluster of co-expressed genes. The global hierarchical classification allows finding various sub-groups of genes, unifies the main variants of RXA algorithms and explores a much larger solution space compared to current solutions based on exhaustive search. Finally, the multi-objective fitness function, which includes accuracy, discriminative power of genes and clusters consistency, as well as specialized variants of genetic operators improve evolutionary convergence and reduce model underfitting. Importantly, patterns in predictive structures are kept comprehensible and may have direct applicability. Experiments carried out on 8 cancer-related gene expression datasets show that the proposed approach allows finding interesting patterns and significantly improves the accuracy of predictions.

This paper discusses the multi-objective evolutionary approach to induction of model trees. The model tree is a particular case of a decision tree designed to solve regression problems. Although the decision tree induction is inherently a multi-objective task, most of the conventional learning algorithms can only deal with a single objective that may possibly aggregate multiple objectives. The goal of this paper is to demonstrate how a set of non-dominated model trees can be obtained using the Global Model Tree (GMT) system. The GMT framework can be used for the evolutionary induction of different types of decision trees, including univariate, oblique or mixed; regression and model trees. Proposed Pareto approach for GMT allows the decision maker to select desired output model according to his preferences on the conflicting objectives. Performed study covers the regression trees and the model trees with two or three objectives that relate to the tree error and the tree comprehensibility. Experimental evaluation discusses the importance of multi-objective components like crowding function and archive elitist selection, using real-life datasets. Finally, the presented multi-objective GMT solution is confronted with competitive regression and model tree inducers.

The paper presents a computational model of magnetic resonance (MR) flow imaging. The model consists of three components. The first component is used to generate complex vascular structures, while the second one provides blood flow characteristics in the generated vascular structures by the lattice Boltzmann method. The third component makes use of the generated vascular structures and flow characteristics to simulate MR flow imaging. To meet computational demands, parallel algorithms are applied in all the components. The proposed approach is verified in three stages. In the first stage, experimental validation is performed by an in vitro phantom. Then, the simulation possibilities of the model are shown. Flow and MR flow imaging in complex vascular structures are presented and evaluated. Finally, the computational performance is tested. Results show that the model is able to reproduce flow behavior in large vascular networks in a relatively short time. Moreover, simulated MR flow images are in accordance with the theoretical considerations and experimental images. The proposed approach is the first such an integrative solution in literature. Moreover, compared to previous works on flow and MR flow imaging, this approach distinguishes itself by its computational efficiency. Such a connection of anatomy, physiology and image formation in a single computer tool could provide an in silico solution to improving our understanding of the processes involved, either considered together or separately.

This paper presents two approaches in parallel modeling of vascular system development in internal organs. In the first approach, new parts of tissue are distributed among processors and each processor is responsible for perfusing its assigned parts of tissue to all vascular trees. Communication between processors is accomplished by passing messages, and therefore, this algorithm is perfectly suited for distributed memory architectures. The second approach is designed for shared memory machines. It parallelizes the perfusion process during which individual processing units perform calculations concerning different vascular trees. The experimental results, performed on a computing cluster and multicore machines, show that both algorithms provide a significant speedup.

The use of quantitative imaging for the characterization of hepatic tumors in magnetic resonance imaging (MRI) can improve the diagnosis and therefore the treatment of these life-threatening tumors. However, image parameters remain difficult to interpret because they result from a mixture of complex processes related to pathophysiology and to acquisition. These processes occur at variable spatial and temporal scales. We propose a multiscale model of liver dynamic contrast-enhanced (DCE) MRI in order to better understand the tumor complexity in images. Our design couples a model of the organ (tissue and vasculature) with a model of the image acquisition. At the macroscopic scale, vascular trees take a prominent place. Regarding the formation of MRI images, we propose a distributed model of parenchymal biodistribution of extracellular contrast agents. Model parameters can be adapted to simulate the tumor development. The sensitivity of the multiscale model of liver DCE-MRI was studied through observations of the influence of two physiological parameters involved in carcinogenesis (arterial flow and capillary permeability) on its outputs (MRI images at arterial and portal phases). Finally, images were simulated for a set of parameters corresponding to the five stages of hepatocarcinogenesis (from regenerative nodules to poorly differentiated HepatoCellular Carcinoma).

In this paper, a model-based approach to medical image analysis is presented. It is aimed at understanding the influence of the physiological (related to tissue) and physical (related to image modality) processes underlying the image content. This methodology is exemplified by modeling first, the liver and its vascular network, and second, the standard computed tomography (CT) scan acquisition. After a brief survey on vascular modeling literature, a new method, aimed at the generation of growing three-dimensional vascular structures perfusing the tissue, is described. A solution is proposed in order to avoid intersections among vessels belonging to arterial and/or venous trees, which are physiologically connected. Then it is shown how the propagation of contrast material leads to simulate time-dependent sequences of enhanced liver CT slices.