Descemet membrane endothelial keratoplasty (DMEK) represents a pivotal advance in corneal transplantation, specifically targeting cases of endothelial dysfunction. Postoperative monitoring of endothelial cell density (ECD) is critical for graft health and patient outcomes. Therefore, this study explores the potential of artificial intelligence to automate corneal endothelial cell counting and density estimation in patients with DMEK using confocal microscopy images. We retrospectively train an encoder-decoder convolutional neural network for cell center detection using clinically annotated patient data. The trained model detects cell centers in images from unseen patients. The image processing techniques then define the region of interest and measure its size, enabling estimation of the density of the cell. The performance of the method is evaluated in DMEK patients at 1–18-month follow-up intervals. Our automatic method achieves a relative error of 23% compared to manual expert measurements. Expert ECD measurements from confocal and specular microscopy vary by 32% on average for the same patients. Our study provides a technical foundation for automated ECD estimation and has a potential to impact post-DMEK care. By enabling frequent ECD monitoring, we can detect graft complications earlier and tailor treatment using longitudinal ECD data.

Training robust image segmentation models, particularly for medical applications like brain lesion segmentation, necessitates large, accurately annotated datasets. However, acquiring such datasets is often expensive and time-consuming. To address this challenge, we propose leveraging generative image models, specifically diffusion models, to synthesize realistic MRI brain images.This study investigates the efficacy of two approaches: utilizing ControlNet and employing a custom-designed diffusion model. We evaluate the impact of these synthetic images on U-Net model training in two scenarios: (i) training on real, artificial, and mixed datasets of equal size, and (ii) incrementally augmenting a small real dataset with increasing amounts of synthetic data. We also benchmark the diffusion models against GAN-based competitors to provide a more comprehensive evaluation of their performance.Our findings demonstrate that both diffusion models effectively generate high-quality brain outperforming GANs in this task. Specifically, synthetic data generated by the diffusion models effectively approximates real brain image data, with minimal impact on segmentation performance, with only a 1.5% performance loss observed. Also, it significantly improves the U-Net’s brain lesion segmentation performance when incorporated into the training process. Finally, our findings demonstrate that diffusion-generated images are more effective than GAN-generated images for improving segmentation model training.This research highlights the potential of diffusion models to overcome data limitations in medical image analysis and pave the way for more accurate and efficient lesion segmentation.

The Fréchet Coefficient (FC) is a reliable and interpretable metric for evaluating generative models, including GANs and diffusion models. Unlike Fréchet Distance, including FID, i.e., its InceptionV3-based version, FC demonstrates stability across feature extractors, robustness to feature dimensionality, and a bounded scale [0–1], enabling transparent and consistent comparisons. Experiments on CIFAR-10, CelebA-HQ, AFHQv2, FFHQ, and ImageNet datasets show FC’s ability to track model quality improvements and distinguish subtle differences across architectures. Its computational efficiency and versatility make it a superior alternative to traditional metrics, positioning FC as a robust tool for the standardized evaluation of generative models across diverse datasets and feature extractors.

Generative adversarial networks (GANs) have shown remarkable capabilities for synthesizing realistic images and movies. However, evaluating the performance of GANs remains a challenging task. Specifically, existing metrics dedicated to this task, such as the Fr & eacute;chet Inception Distance (FID), lack interpretability since they provide scores that are not bound to any range. This paper introduces the Fr & eacute;chet Coefficient (FC), the novel metric that addresses the challenge by providing a clear performance score between 0 and 1, thus making interpreting and comparing results easier. Also, FC can use any convolutional neural network as a feature extractor, offering flexibility and potential for customization. We evaluate the performance of FC and benchmark it against FID on five diverse image datasets within the image-to-image translation framework. These datasets include medical, natural scene, and face images where GANs are tasked with synthesizing images from semantic segmentation maps. We also test FC's performance under various image distortions. Experimental results demonstrate that FC is a reliable metric for evaluating GAN performance. It consistently outperforms FID regarding interpretability, making it a valuable tool for researchers and practitioners working with GANs.

This paper proposes a novel approach to assessing the quality of 3D Electrical Capacitance Tomography (ECT) images. Such images are typically represented as irregular graphs. Thus, image quality metrics typically used with raster images do not straightforwardly apply to them. However, given the recent advancements in Graph Convolutional Neural Networks (GCNs) for improving ECT image reconstruction, reliable Quality Assessment methods are essential for comparing the performance of different GCN models. To address this need, this paper applied some existing image quality and similarity assessment methods designed for raster images to the graph-based representation of 3D ECT images. Specifically, attention was paid to the Peak Signal-to-Noise Ratio (PSNR), the Structural Similarity Index Measure (SSIM), and measures based on image histograms. The proposed adaptations resulted in the development of tailored Graph Quality Assessment (GQA) techniques specifically designed for the graph-based nature of ECT images. The proposed GQA techniques were validated on 1042 phantoms and their corresponding Low-Quality (LQ) and High-Quality (HQ) reconstructions through a robust GQA benchmarking system, enabling a systematic comparison of various GQA methods. The evaluation of the proposed methods’ performances across this diverse dataset, by analyzing overall trends and specific case studies, is presented and discussed. Finally, we present our conclusions regarding the effectiveness of the proposed GQA methods, and we identify the most promising approach for assessing the quality of graph-based ECT images.

The structure and features of resin ducts provide valuable information about environmental conditions accompanying the growth of trees in the genus Pinus. Therefore analysis of resin duct characteristics has been an increasingly common measurement in dendrochronology. However, the measurement is tedious and time-consuming since it requires thousands of ducts to be manually marked in an image of an enlarged wood surface. Although tools exist to automate some stages of this process, no tool exists to automatically recognize and analyze the resin ducts and standardize them with the tree rings they belong to. This study proposes a new fully automatic pipeline that quantifies the properties of resin ducts in terms of the tree ring area to which they belong. A convolutional neural network underlays the pipeline to detect resin ducts and tree-ring boundaries. Also, a region merging procedure is used to identify connected components corresponding to successive rings. Corresponding ducts and rings are next related to each other. The pipeline was tested on 74 wood images representing five Pinus species. Over 8000 tree-ring boundaries and almost 25,000 resin ducts were analyzed. The proposed method detects resin ducts with a sensitivity of 0.85 and precision of 0.76. The corresponding scores for tree-ring boundary detection are 0.92 and 0.99, respectively.

This paper considers the problem of corneal endothelium image segmentation using a method that combines a CNN model with a watershed transform. Specifically, first CNN predicts cell bodies, edges, and centers. Next, cell centers are used as markers that guide the watershed transform performed concerning the cell edge probability maps inferred by the CNN to outline cell edges. Different variants of the method are considered. Specifically, a downscaled U-Net is compared with the Attention U-Net in the image-to-image and sliding window setup. Results show that using a marker-driven watershed transform to post-process cell edge probability maps allows for replacing the sliding window setup with an image-to-image setup, reducing prediction time while maintaining similar or better segmentation accuracy. Also, when used as a backbone, Attention U-Net outperforms classical U-Net in determining cell morphometric parameters with high accuracy.

Objective. Current solutions for corneal endothelial image segmentation use convolutional neural networks. However, their potential is not exploited due to the scarcity of labeled corneal endothelial data caused by an expensive cell delineation process. Therefore, this work proposes synthesizing cell edges and corresponding images using generative adversarial neural networks. To our knowledge, such work has not yet been reported in the context of corneal endothelial image segmentation. Methods. A two-step pipeline for synthesizing training patches is proposed. First, a custom mosaic generator creates a grid that mimics a binary map of endothelial cell edges. The synthetic cell edges are then input to the generative adversarial neural network, which is trained to generate corneal endothelial image patches from the corresponding edge labels. Finally, pairs of synthetic patches are used to train the patch-based U-Net model for corneal endothelial image segmentation. Results. Experiments performed on three datasets of corneal endothelial images show that using purely synthetic data for U-Net training allows image segmentation with comparable accuracy to that obtained when using original images annotated by experts. Depending on the dataset, the segmentation quality decreases only from 1% to 4%. Conclusions: Our solution provides a cost-effective source of diverse training data for corneal endothelial image segmentation. Significance. Due to the simple graphical user interface wrapping the proposed solution, many users can easily obtain unlimited training data for corneal endothelial image segmentation and use it in various scenarios.

Abstract The correct identification of timber species is a complicated task for the wood industry and government institutions regulating the different laws that ensure legal and transparent commerce. Currently, experts perform this process using the organoleptic characteristics of the wood. However, the methodology used is time-consuming and limited to environmental conditions. Moreover, it has a scalability issue since acquiring this specific knowledge and experience has a slow learning curve. On the other hand, deep learning models have evolved as possible solutions for process automation. Therefore, this paper explores convolutional neural network models suited to run on edge devices. The present study created a database with 25k images of 25 timber species from the Peruvian Amazon. We trained-validated multiple lightweight models (less than 5M). The experiments were made using a repeated stratified k-fold cross-validation approach to estimate the performance of the classifiers. The experiments show that the best model has an F1 score metric of 99.90\% and 58ms latency using 561k parameters. Furthermore, the created model showed an excellent ability to identify species, opening up space for future integration with mobile applications, which helps minimize the time spent and the identification errors on timber identification carried out by experts on control points.

The problem of image segmentation is one of the most significant ones in computer vision. Recently, deep-learning methods have dominated state-of-the-art solutions that automatically or interactively divide an image into subregions. However, the limitation of deep-learning approaches is that they require a substantial amount of training data, which is costly to prepare. An alternative solution is semi-supervised image segmentation. It requires rough denotations to define constraints that are next generalized to precisely delimit relevant image regions without using train examples. Among semi-supervised strategies for image segmentation, the leading are graph-based techniques that define image segmentation as a result of pixel or region affinity graph partitioning. This paper revisits the problem of graph-based image segmentation. It approaches the problem as semi-supervised node classification in the SLIC superpixels region adjacency graph using a graph convolutional network (GCN). The performance of both spectral and spatial graph convolution operators is considered, represented by Chebyshev convolution operator and GraphSAGE respectively. The results of the proposed method applied to binary and multi-label segmentation are presented, numerically assessed, and analyzed. In its best variant, the proposed method scored the average DICE of 0.86 in the binary segmentation task and 0.79 in the multi-label segmentation task. Comparison with state-of-the-art graph-based methods, including Random Walker and GrabCut, shows that graph convolutional networks can represent an attractive alternative to the existing solutions to graph-based semi-supervised image segmentation.

Quantitive information about corneal endothelium cells’ morphometry is vital for assessing cornea pathologies. Nevertheless, in clinical, everyday routine dominates qualitative assessment based on visual inspection of the microscopy images. Although several systems exist for automatic segmentation of corneal endothelial cells, they exhibit certain limitations. The main one is sensitivity to low contrast and uneven illumination, resulting in over-segmentation. Subsequently, image segmentation results often require manual editing of missing or false cell edges. Therefore, this paper further investigates the problem of corneal endothelium cell segmentation. A fully automatic pipeline is proposed that incorporates the watershed algorithm for marker-driven segmentation of corneal endothelial cells and an encoder-decoder convolutional neural network trained in a sliding window set up to predict the probability of cell centers (markers) and cell borders. The predicted markers are used for watershed segmentation of edge probability maps outputted by a neural network. The proposed method's performance on a heterogeneous dataset comprising four publicly available corneal endothelium image datasets is analyzed. The performance of three convolutional neural network models (i.e., U-Net, SegNet, and W-Net) incorporated in the proposed pipeline is examined. The results of the proposed pipeline are analyzed and compared to the state-of-the-art competitor. The obtained results are promising. Regardless of the convolutional neural model incorporated into the proposed pipeline, it notably outperforms the competitor. The proposed method scored 97.72% of cell detection accuracy, compared to 87.38% achieved by the competitor. The advantage of the introduced method is also apparent for cell size, DICE coefficient, and Modified Hausdorff distance.

This paper tackles the problem of automatic tree species identification from scanned images of wood cores. A convolutional neural network with residual connections is proposed to perform this task. The model is applied to consecutive image patches following the sliding window strategy to recognize a patch central pixel's membership. It then decides about the resulting tree species via a majority voting. The model's performance was assessed concerning a dataset of 312 wood core images representing 14 European tree species, including both conifer and angiosperm (ring-porous and diffuse-porous) wood. Two tasks were considered, including wood patch classification and wood core classification. In these tasks, the proposed model correctly recognized species of almost 93% the wood image patches and 98.7% of wood core images. It also outperformed the state-of-the-art convolutional neural network-based competitor by 9% and 3%, respectively. The influence of the model's parameters and training set-up on its performance is analyzed in the manuscript to ensure the highest recognition rates of wood species. The source code of the proposed method is released together with the corresponding image dataset to facilitate the reproduction of results.

Accurate liver vessel segmentation is of crucial importance for the clinical diagnosis and treatment of many hepatic diseases. Recent state-of-the-art methods for liver vessel reconstruction mostly utilize deep learning methods, namely, the U-Net model and its variants. However, to the best of our knowledge, no comparative evaluation has been proposed to compare these approaches in the liver vessel segmentation task. Moreover, most research works do not consider the liver volume segmentation as a preprocessing step, in order to keep only inner hepatic vessels, for Couinaud representation for instance. For these reasons, in this work, we propose using accurate Dense U-Net liver segmentation and conducting a comparison between 3D U-Net models inside the obtained volumes. More precisely, 3D U-Net, Dense U-Net, and MultiRes U-Net are pitted against each other in the vessel segmentation task on the IRCAD dataset. For each model, three alternative setups that allow adapting the selected CNN architectures to volumetric data are tested, namely, full 3D, slab-based, and box-based setups are considered. The results showed that the most accurate setup is the full 3D process, providing the highest Dice for most of the considered models. However, concerning the particular models, the slab-based MultiRes U-Net provided the best score. With our accurate vessel segmentations, several medical applications can be investigated, such as automatic and personalized Couinaud zoning of the liver.

BACKGROUND AND OBJECTIVE:Blood flow detection through the retinal vessels is an essential step in diagnosing several eye diseases. It manifests itself as changes in vessel diameter in consecutive phases of blood flow. Previous studies focused mainly on determining retinal vessel diameter by the manual demarcation of vessel edges, which is time-consuming. As a result, only a few selected vessels were considered, which is not reliable. Such measurements are also prone to human errors and operator subjectivity, which additionally decreases their reliability. For these reasons, this paper proposes an automated pipeline to analyze the blood flow through retinal vessels.METHODS:Convolutional neural networks were used for optic disc and vessel detection and full width at half maximum analysis used for vessel width assessment at the subpixel level. Measurements of the diameter were performed for five phases of the blood flow to all vessels at a particular distance derived from the optic disc size. We tested the approach on fundus images of five patients, with both eyes examined in each participant. The threshold for the detections of blood flow was when the retinal diameter vessel measurements were above the camera's resolution as compared among all 5 phases of blood flow.RESULTS:A total of 205 large caliber blood vessels were analyzed with blood flow detected in 18 retinal blood vessels. Conclusions Average vessel diameters derived from manual and automatic measurements differed on average by 4.96%. Average relative errors for single vessel measurements along the vessels range from 4.21 to 11.85%, with a global average at the level of 8%. Therefore, the measurements can be considered as accurate and in a high agreement between the expert and algorithm.

Three dimensional Electrical Capacitance Tomography (3D ECT) is an inexpensive tool for diagnosing non-conductive components of industrial processes. Although relatively mature, it still requires much work to improve its inverse nature of imaging capability. In particular, high resolution 3D ECT image reconstruction is very time-consuming and computationally heavy, and the best-known 3D ECT image reconstruction techniques have already reached their limits. Thus, there is a strong need to change a direction towards modern computational intelligence solutions. Therefore, this work proposes using graph convolutional networks (GCN) to raise the 3D ECT image quality. Mainly, it takes advantage of GCN's ability to effectively use specific geometrical relationships hidden in the finite modeling unstructured grids commonly used to build 3D ECT images. These relationships are first encoded by a graph representing an ECT volumetric finite element grid. A GCN is next trained in a graph-to-graph framework with pairs of graphs representing high-quality nonlinear image reconstruction results as input and a simulated phantom as output. As a result, a trained GCN model fed with lower resolution 3D ECT image enhances its quality and spatial resolution. Tomographic image quality and resolution enhancement was evaluated using normalized mean square error and Pearson correlation coefficient, which improved by 35.5% and 3.74%, respectively.

This paper considers the problem of airway tree matching in 3D images. An algorithm dedicated to chest scans of one patient obtained at different breathing stages is proposed. It assumes that the airway trees were already segmented from the 3D images. The method gradually transforms the moving tree to match a fixed tree, recursively rotating consecutive branches and their subbranches. The experiments were performed on 3D CT datasets of three patients, with images representing lungs in successive breathing stages. The assessment was made via the DICE coefficient between the fixed and the moving tree. Due to transformation, the coefficient increased by up to 30

This paper considers the problem of pig automatic segmentation from infrared top view images of a pen. Particularly, an algorithm for accurate delineation of pig’s contour is presented. The method consists of two main steps. In the first step, a rough contour is determined using standard image processing methods. Next, the initial contour is gradually deformed so that it reflects the actual contour of the pig as much as possible. This effect is obtained by attracting initial contour points to the nearest local gradient peaks. In the last step, the contour is refined and smoothed by removing loops. This step incorporates analysis of the angles between contour segments passing through the consecutive contour points. Results of the proposed approach for sample infrared images of pigs in a pen are presented and discussed. They reveal that the method performs reasonably well with the average DICE score exceeding the level of 0.97 and the average Jaccard index above 0.95.

Varves – annual sediment layers, common in glacial lakes – are an important source of paleoclimate information. Manually recording their occurrence, typically by visual inspection, can be both time-consuming and prone to error, leading to several attempts in recent years to at least partially computerize the process. However, existing computerized methods of varve detection still require moderate to large amounts of user interaction — they are semi-automated, rather than fully automated. In light of that, this paper is a step towards fully automatic detection of varves. The presented program, DeepVarveNet - a glacial varve detector built on a convolutional neural network, is designed to automatically delineate annual layers in such lacustrine sediment in digital images of photographed sediment cores. To the best of the authors’ knowledge, this is the first approach that applies a convolutional neural network to this task. The performance of DeepVarveNet was assessed on a data set comprising images from seven sediment coring sites, of varying sedimentological properties. They represent three northeast U.S. glacial paleolakes, and glacial paleolake Ojibway. Our testing set contained 1415 identified varves, on which DeepVarveNet demonstrated sensitivity at a level of 0.986 and precision equal to 0.834, exceeding that of BMPix and ANFIS, the existing semi-automated varve identifiers.