Abstract Brain tumors pose a serious threat in our modern society, with a clear increase in global cases each year. Therefore, developing robust solutions that could automatically and reliably detect brain tumors in their early stages is of utmost importance. In our paper, we revisit the problem of building performant ensembles for clinical usage by maximizing the diversity of the member models during the training procedure. We present an improved, more robust, extended version of our framework and propose solutions that could be integrated into a Computer-Aided Diagnosis system to accurately classify some of the most common types of brain tumors: meningioma, glioma, and pituitary tumors. We show that the new framework based on the histogram loss can be seen as a natural extension of the former approach, as it also calculates the inner products of the latent vectors produced by each member to measure similarity, but at the same time, it also makes it possible to capture more complex patterns. We also present several variants of our framework to incorporate member models with varying dimensional feature vectors and to cope with imbalanced datasets. We evaluate our solutions on a clinically tested dataset of 3,064 T1-weighted contrast-enhanced magnetic resonance images and show that they greatly outperform other state-of-the-art approaches and the base architectures as well, achieving over 92% accuracy, 92% macro and weighted precision, 91% macro and 92% weighted $$F_{1}$$ F 1 score, and over 90% macro and 92% weighted sensitivity.

This research investigates how well different deep learning architectures classify medical imaging data, with a particular emphasis on identifying schizophrenia. Six models were assessed: a 4D ResNet architecture, two CNNs, a CNN-LSTM hybrid, EfficientNetV2, and MobileNetV3. Using the COBRE dataset, the study used 5-fold cross-validation to assess these models’ performance. Besides evaluating deep learning architectures, this work includes a pretreatment pipeline for fMRI data and exploratory data analysis. Data is arranged for effective administration, and dimensionality is reduced using methods like PCA. A test accuracy of 94.7% was accomplished by the first CNN model, and 99.75% by the enhanced CNN model. The CNN-LSTM hybrid, in particular, demonstrated remarkable performance with a test accuracy of 99.74%. EfficientNetV2 and MobileNetV3, on the other hand, had accuracies that were 93.23% and 63.41%, respectively, lower. At 60.00% test accuracy, the 4D ResNet model produced the least desirable outcome. These results highlight how crucial it is to choose the right architectures for medical picture classification tasks, especially in light of resource constraints. Taking into consideration hardware limitations, CNN-based models, in particular the CNN-LSTM hybrid and the second CNN model, hold potential for additional research in this area.

Abstract We introduce new tools to support finding efficient graph embedding techniques for graph databases and to process their outputs using deep learning for classification scenarios. Accordingly, we investigate the possibility of creating an ensemble of different graph embedding methods to raise accuracy and present an interconnected neural network-based ensemble to increase the efficiency of the member classification algorithms. We also introduce a new convolutional neural network-based architecture that can be generally proposed to process vectorized graph data provided by various graph embedding methods and compare it with other architectures in the literature to show the competitiveness of our approach. We also exhibit a statistical-based inhomogeneity level estimation procedure to select the optimal embedding for a given graph database efficiently. The efficiency of our framework is exhaustively tested using several publicly available graph datasets and numerous state-of-the-art graph embedding techniques. Our experimental results for classification tasks have proved the competitiveness of our approach by outperforming the state-of-the-art frameworks.

We present a novel ensemble-based method for the accurate and reliable classification of some of the most common types of brain tumors: meningioma, glioma, and pituitary tumors. We show that using the proposed method, we can re-use the same neural architecture and build diverse ensembles that surpass the performance of the original architectures and could potentially be used as a part of a Computer Aided Diagnosis system. In this paper, we demonstrate that we can treat the features extracted by the convolutional layers of the members as elements of a latent vector space and use cosine similarity as a measure to calculate their similarities. By using a custom loss function and regularizing the members based on their internal similarities, we show that we can train diverse ensemble models that can generalize well to unseen data and classify brain tumors with high accuracy, sensitivity, and precision. We experiment with multiple state-of-the-art CNN models that are frequently used in the clinical setting and show that we can surpass all of them in terms of performance, achieving over 91% accuracy and precision and over 90% sensitivity and F1-score.

In this study, we attempted to develop a method for accelerating parameter optimization of an object detector ensemble over large image datasets by using simulated annealing. We propose a novel sampling-based evaluation method that considers the minimum portion of the dataset required in each iteration to maintain solution quality. This approach can be considered a noisy evaluation of the energy. The sample sizes required during the search process are theoretically determined by adapting the convergence results for noisy evaluation. To determine applicability, we prepared and optimized two ensembles for diabetic retinopathy pre-screening based on microaneurysm detection with convolutional neural network-based and traditional object detectors. Our experimental results indicate that the proposed sampling-based evaluation method substantially reduced the computational time required for optimizing the parameters of the ensembles while preserving solution quality.

Endoscope-based surgery has several beneficial effects regarding the rehabilitation of the patients, but has some drawbacks causing difficulties to medical experts, on the contrary. The main disadvantage is that the tactile information is loosed to expert who takes the surgical intervention. There are some organs (e.g. ureters and arteries) in the human body which have similar visual appearances, so the differentiation of them based on only visual expression via endoscopy is a challenging task to the medical experts. To support keyhole-surgery using state-of-the-art image processing solutions, we have developed a semi-automatic software which can distinguish ureters from arteries by a dedicated convolutional neural network (CNN). We have trained the CNN on 2000 images acquired during endoscopic surgery and tested on 500 test ones. 94.2% accuracy has been achieved in this two-classes classification task regarding a binary error function.

The presence of atypical (irregular) pigment networks can be a symptom of melanoma malignum in skin lesions, thus, their proper recognition is a critical task. For object classification problems, the application of deep convolutional neural nets (CNN) receives priority attention nowadays for their high recognition rate. The descriptive features found by CNNs are more hidden than the classically applied ones for texture recognition. In this paper, we investigate whether CNN features outperform the classical texture descriptors in the classification of typical/atypical pigment network. Beyond performing this analysis, we have also found that the aggregation of CNN and classical features within a joint classification framework had a superior performance. Specifically, the union of the CNN and classical feature sets leads to a much higher stability in classification performance for various classifiers. As for quantitative figures, we have reached 90.44% recognition accuracy using a specific subset of this combined feature set obtained by linear forward feature selection and using a Bayes Net as classifier.

In a recent work, we have proposed a novel way to approximate point sets with grids using the LLL algorithm, which operates in polynomial time. Now, we show how this approach can be applied to pattern recognition purposes with interpreting the rate of approximation as a new feature for regularity measurement. Our practical problem is the characterization of pigment networks in skin lesions. For this task we also introduce a novel image processing method for the extraction of the pigment network. Then, we show how our grid approximation framework can be applied with specializing it for the recognition of hexagonal patterns. The classification performance of our approach for the pigment network characterization problem is measured on a database annotated by a clinical expert. Throughout the paper we address several practical issues that may help to apply our general framework to other practical tasks, as well.

Ensemble-based systems are primarily analyzed on how the accuracy of the ensemble depends on that of its members. In this paper, we extend this model with adding a natural constraint regarding a time limit within which the ensemble should make the decision. For this aim, we consider both the execution time and accuracy of each member. Then, we solve the problem on how to find the most accurate ensemble, where the sum of the execution times of its members remains below the limit. As a decision rule, we analyze a majority voting-based one generalized to be applicable in single object detection scenarios. The optimization task leads to a non-separable Knapsack problem, which is addressed using stochastic considerations. The proposed methodology is also validated experimentally for the localization of the optic disc in retinal images.

Medical image processing plays an important role in the automatic detection of certain eye diseases, disease documentation, treatment monitoring and educational purposes. The continuous improvement of the hardware of mobile phones gives us the ability to capture images like this. Interesting question can be: are these images suitable for automatic image processing, or more specifically, how much information can be obtained from this kind of images? In this paper, we would like to investigate whether a mobile device suitable to use as a medical diagnostic tool. We present a comparison between images made by a mobile device and a clinical eye examination one. We perform optic disc and optic cup detection from which we calculate cup to disc ratio. Our results show that images acquired by mobile phone are suitable for automatic image processing. Beyond these, a database of images were set up to allow researchers to test their algorithms and methods for image processing.

In this paper, we show how the optimal setting of an ensemble-based lesion detector can improve the performance of the system. That is, instead of the individually optimal parameter settings of the member algorithms, we determine such a setting that maximizes the accuracy of the ensemble. Our case study is an ensemble composed for microaneurysm detection on retinal images that is currently ranked as first in an online competition. The determination of the optimal parameter setting is a very time consuming problem, especially in fusion-based systems. Thus, we suggest several speedups in a stochastic optimum search framework regarding the evaluation of the corresponding special energy function. In other words, we also suggest some modifications and tuning of stochastic search approaches for ensemble-based image detectors evaluated on test databases.

Generating ensembles from multiple individual classifiers is a popular approach to raise the accuracy of the decision. As a rule for decision making, majority voting is a usually applied model. In this paper, we generalize classical majority voting by incorporating probability terms pn,k to constrain the basic framework. These terms control whether a correct or false decision is made if k correct votes are present among the total number of n. This generalization is motivated by object detection problems, where the members of the ensemble are image processing algorithms giving their votes as pixels in the image domain. In this scenario, the terms pn,k can be specialized by a geometric constraint. Namely, the votes should fall inside a region matching the size and shape of the object to vote together. We give several theoretical results in this new model for both dependent and independent classifiers, whose individual accuracies may also differ. As a real world example, we present our ensemble-based system developed for the detection of the optic disc in retinal images. For this problem, experimental results are shown to demonstrate the characterization capability of this system. We also investigate how the generalized model can help us to improve an ensemble with extending it by adding a new algorithm.

A method for the automatic detection of microaneurysms (MAs) in color retinal images is proposed in this paper. The recognition of MAs is an essential step in the diagnosis and grading of diabetic retinopathy. The proposed method realizes MA detection through the analysis of directional cross-section profiles centered on the local maximum pixels of the preprocessed image. Peak detection is applied on each profile, and a set of attributes regarding the size, height, and shape of the peak are calculated subsequently. The statistical measures of these attribute values as the orientation of the cross-section changes constitute the feature set that is used in a naïve Bayes classification to exclude spurious candidates. We give a formula for the final score of the remaining candidates, which can be thresholded further for a binary output. The proposed method has been tested in the Retinopathy Online Challenge, where it proved to be competitive with the state-of-the-art approaches. We also present the experimental results for a private image set using the same classifier setup.

Nowadays diabetic retinopathy is one of the most common reasons of blindness in the world. Exudates are the primary sign of this disease so the proper detection of these lesions is an essential task in an automatic screening system. In this paper, we propose a method for exudate detection which performs with high accuracy. First, we identify possible regions containing exudates using grayscale morphology. Then, we extract more than 50 descriptors for each candidate pixel to classify them. We analyzed the information content of the descriptors and selected the most relevant ones. The selected features are used to train a boosted naïve Bayes classifier. We tested this approach on the publicly available DiaretDB color fundus image database, where the proposed detector outperformed the state-of-the-art ones regarding the FScore.

In this paper, we propose a novel approach to improve microaneurysm detection in color fundus images by clustering image databases since they usually contain images with different characteristics. Thus, a parameter setting of an algorithm determined for a database is not necessarily optimal on another one. To overcome this problem, we determine clusters of retinal images coming from different sources. In other words, we consider individual image characteristics instead of databases in a detection problem. We select 19 similarity measures to calculate image differences, and apply k-means clustering to obtain the clusters. For each cluster, an optimal parameter setting is determined for the same microaneurysm detector. We tested our approach on a publicly available database, where the performance of a state-of-the-art microaneurysm detector is successfully increased by the proposed method.

The key to the early detection of diabetic retinopathy is to recognize microaneurysms in the fundus of the eye in time. Reliable detection of such lesions is still an open issue in medical image processing. We propose a framework which assembles several candidate extractors and preprocessing methods to strengthen the detection accuracy of the individual approaches. We use a simulated annealing based search approach to select an optimal combination from the available methods. The proposed method has proved its superiority over the individual algorithms in an online competition dedicated to the objective comparison of microaneurysm detectors.

The proper segmentation of the vascular system of the retina has a very important role in automatic screening systems. Its detection helps the localization of other anatomical parts and also the detection of possible vascular disorders. State-of-the-art machine learning algorithms are reported to have good performance in this field. However, with the spatial resolution of the fundus images growing, it is necessary to decrease the number of training pixels to save computations. In this paper, we investigate several subsampling strategies with the motivation to find the best segmentation results with involving fewer pixels into the analyses. Besides checking the computational advantages, we demonstrate how the segmentation accuracy drops with the level of subsampling.

The use of an alphabet of line segments to compose a curve is a possible approach for curve data compression. Many approaches are developed with the drawback that they can process simple curves only. Curves having more sophisticated topology with self-intersections can be handled by methods considering recursive decomposition of the canvas containing the curve. In this paper, we propose a graph theory-based algorithm for tracing the curve directly to eliminate the decomposition needs. This approach obviously improves the compression performance, as longer line segments can be used. We tune our method further by selecting optimal turns at junctions during tracing the curve. We assign a polygon approximation to the curve which consists of letters coming from an alphabet of line segments. We also discuss how other application fields can take advantage of the provided curve description scheme.

Active contour (snake) approaches have been proved to be efficient tools to extract object boundary precisely. One drawback of these methods is that state-of-the-art snake algorithms are usually computationally expensive. To overcome this difficulty, one approach is to reduce spatial resolution to make snake iteration faster. The disadvantage of this method is that the reduction of resolution ignores image content which may lead to losing useful image data. In this paper we propose a snake iteration method on quadtree representations derived from the image. In this way we can reduce resolution in a special and adaptive way. Once the iteration is finished on the sparse quadtree grid, final adjustments can be made by performing pixelwise snake iteration steps.

The use of an alphabet of line segments to compose a curve is a possible approach for curve data compression. An existing state-of-the-art method considers a quadtree decomposition of the curve to perform the substitution of the curve parts from the alphabet of line segments. In this paper, we propose a graph theory based algorithm for tracing the curve directly to eliminate the quadtree decomposition needs. This approach obviously improves the compression efficiency, as longer line segments can be used. We tune our method further by selecting optimal turns at junctions during tracing the curve. We also discuss briefly how other application fields can take advantage of the presented approach.

This paper proposes a novel algorithm for an optimal reduction of object description for object matching purposes. Our aim is to decrease the computation needs by considering simplified objects, thus reducing the number of pixels involved in the matching process. We develop the appropriate theoretical background based on centroidal Voronoi tessellations. Its use within the chamfer matching framework is also discussed. We present experimental results regarding the performance of this approach for 2-D contour and region-like object matching. As a special case, we investigate how the snake based representation of target objects can be employed in chamfer matching. The experimental results concern the use of object part matching for recognizing humans and show how the proposed simplification leads to valid replacements of the original templates.

In this paper we present a novel method on how to take advantage of the snake representation of target objects, when doing chamfer matching for detection/recognition purposes. In this case several time-consuming steps of classic chamfer matching approaches can be simplified. Moreover, we investigate the possibility of involving fewer pixels from both the target and template object to speed up computations. We introduce an optimization method for such an object reduction, which is valid also in the general application scheme of chamfer matching. Finally, we present our experimental results regarding human body detection

In this paper we present some methods for indexing and segmenting colour images. The proposed procedures are based on well-known algorithms, but now we use digital distance functions generated by neighbourhood sequences to measure distance between colours. The application of such distance functions is quite natural and descriptive, since the colour coordinates of the pixels are nonnegative integers. An additional interesting property of neighbourhood sequences is that they do not generate metric in general, so we can obtain many distance functions in this way. We describe our methods for RGB images in detail, but other image representations also could be considered. Moreover, the proposed methods can be applied in arbitrary dimension without any difficulties.