The efficacy of deep learning in the analysis of biomedical images and its integration of advanced computational techniques into diagnostic workflows has long been studied in many prominent works, contributing to more precise and timely interventions for improved patient outcomes. Synchrotron X-ray tomographic imaging has recently been used for endomyocardial biopsy sample scanning to assess acute cellular rejection after heart transplantation, which greatly increases resolution and clarity of acquired samples compared to other tomographic imaging techniques.The primary objectives of this work are: first, GRAFT-XPCI, a novel classification dataset consisting of synchrotron X-ray tomographic images obtained from heart transplant recipients at varying stages of rejection is proposed. The primary purpose of this dataset is to increase understanding of X-PCI image processing and to enable the training of deep learning models to automate the identification and quantification of subtle alterations in the transplanted heart associated with rejection. Secondly, evaluation protocols are proposed that ensure a fair comparison of different methods on the proposed dataset. Thirdly, a comparison of various existing deep learning models for early rejection detection was conducted. Finally, a set of pre-trained ViT-based foundation models that can be used for further research in the field are provided.

Hierarchical semantic structures are essential in human recognition and understanding of the world. We can recognize shared features between different entities across multiple semantic levels. However, existing deep metric learning methods primarily focus on finding discriminative features between classes on a single semantic level, neglecting some coarser features they share with other classes. This work proposes a method that arranges visual features from multiple semantic levels in a coarse-to-fine manner within a single feature space. Our approach allows us to create scalable embeddings from a single feature vector that can efficiently discriminate on multiple semantic levels and has a clear geometric interpretation. We evaluate our method on three hierarchical image retrieval datasets: DyML-Vehicle, DyML-Animal, and DyML-Product, achieving better or on-par results compared to state-of-the-art methods without compromising performance on the fine-grained level or being biased towards any other semantic level. Finally, we show that our method leads to a more intuitive and better-organized feature space.

Brightness adjustment is an important part of image enhancement, but some of the best brightness adjustment methods in terms of result quality are too complex to run in real-time and many are not suitable for hardware implementation. In this paper a fast learning-based and hardware-friendly method for local brightness adjustment is proposed that in real-time obtains results of higher quality than some of the best methods. The results are presented and discussed. The source code is at http://www.fer.unizg.hr/ipg/resources/color constancy/.

Achieving color invariance to illumination is known as color constancy and it is implemented in most digital cameras. There are statistics-based and learning-based computational color constancy methods and the latter ones are known to be more accurate. For a given image these methods extract some features and since for some methods calculating and processing these features can be computationally demanding, this renders such methods slow and very often impractical for hardware implementation. In this paper simple, yet very powerful features for color constancy based on the red chromaticity are presented. A new color constancy method is proposed and it is demonstrated how an existing one can be simplified. In both cases state-of-the-art results are achieved. The results are presented and discussed and the source code is available at http://www.fer.unizg.hr/ipg/resources/color constancy/.

The 2015 edition of the International Symposium on Image and Signal Processing and Analysis (ISPA 2015) is the ninth in the series of biennial research meetings and it follows the successful ISPA 2013 meeting held in Trieste, Italy. ISPA 2015 is held in Zagreb, the capital city of the Republic of Croatia.

Diabetic retinopathy is one of the leading causes of preventable blindness in the developed world. Early diagnosis of diabetic retinopathy enables timely treatment and in order to achieve it a major effort will have to be invested into screening programs and especially into automated screening programs. Detection of exudates is very important for early diagnosis of diabetic retinopathy. Deep neural networks have proven to be a very promising machine learning technique, and have shown excellent results in different compute vision problems. In this paper we show that convolutional neural networks can be effectively used in order to detect exudates in color fundus photographs.

Having images look the same regardless of the scene illumination is a desirable feature called color constancy. In this paper the Color Cat (CC), a novel fast and accurate learning-based method for achieving computational color constancy is proposed. It learns and then uses the relationship between transformed color histograms and the regularity in the possible illumination colors. The proposed method is tested on a publicly available color constancy dataset and it is shown to outperform most of the other color constancy methods in terms of accuracy and computation cost. The results are presented and discussed. The source code is available at http://www.fer.unizg.hr/ipg/resources/color_constancy/.

In this paper an improvement of the white patch method, a color constancy algorithm, is proposed. The improved method is tested on several benchmark databases and it is shown to outperform the baseline white patch method in terms of accuracy. On the benchmark database it also outperforms most of the other methods and its great execution speed makes it suitable for hardware implementation. The results are presented and discussed and the source code is available at http://www.fer.unizg.hr/ipg/resources/color constancy/.

In this letter, Light Random Sprays Retinex (LRSR), an improvement of the Random Sprays Retinex (RSR) algorithm is proposed. RSR is a white balancing algorithm for achieving local color constancy and image enhancement by using random sprays of the same size. The main problem of the original RSR is that the lower the number and size of the sprays, the greater the noise in the resulting image, which means that the number and size of sprays have to be relatively high in order to reduce the noise leading to a higher computation cost. The proposed improved algorithm is based on a new method to remove the noise in the resulting image thereby allowing only one spray of a smaller size to be used resulting in lower computation cost. By using interpolation the computation cost is reduced even further without a noticeable perceptual difference. The improvement is tested on a public database and is shown to outperform the original RSR in image quality and computation cost. The source code is available at http://www.fer.unizg.hr/ipg/resources/color_constancy/.

Recently, all the craters from the major currently available manually assembled catalogs have been merged into the catalog with 57 633 known Martian impact craters. This paper presents a new crater detection algorithm (CDA) for the search of still uncataloged impact craters. The CDA is based on fuzzy edge detectors and Radon/Hough transform and utilizes digital topography data instead of image data. The critical parts of the method providing increased accuracy are as follows: 1) gradient-value/orientation-based techniques; 2) automated morphometry measurements of depth/diameter ratio, circularity, topographic cross-profile, rim, central peak, and radial range where the crater is preserved; 3) circularity analysis of votes in parameter space; 4) slip tuning of detected craters' parameters; and 5) calibration which partially compensates differences in morphology between small and large craters. Using the framework for the evaluation of CDAs, in comparison with prior work, the proposed detector shows the following: 1) significantly larger area under the free-response receiver operating characteristics (AUROC) and 2) significantly larger number of correct detections. Using the Mars Orbiter Laser Altimeter data as input, the CDA proposed numerous candidates for GT-57633 catalog extension. After the manual survey of all proposed craters and rejection of false detections, 57 592 impact craters were confirmed as correct detections. The accompanying result to the CDA is a new GT-115225 catalog.

PURPOSE: Demonstration of a technique for three-dimensional (3-D) assessment of tracheal-stenoses, regarding site, length and degree, based on spiral computed tomography (S-CT). PATIENTS AND METHODS: S-CT scanning and automated segmentation of the laryngo-tracheal tract (LTT) was followed by the extraction of the LTT medial axis using a skeletonization algorithm. Orthogonal to the medial axis the LTT 3-D cross-sectional profile was computed and presented as line charts, where degree and length was obtained. Values for both parameters were compared between 36 patients and 18 normal controls separately. Accuracy and precision was derived from 17 phantom studies. RESULTS: Average degree and length of tracheal stenoses was found to be 60.5% and 4.32 cm in patients compared with minor caliber changes of 8.8% and 2.31 cm in normal controls (p << 0.0001). For the phantoms an excellent correlation between the true and computed 3-D cross-sectional profile was found (p << 0.005) and an accuracy for length and degree measurements of 2.14 mm and 2.53% respectively could be determined. The corresponding figures for the precision were found to be 0.92 mm and 2.56%. CONCLUSION: LTT 3-D cross-sectional profiles permit objective, accurate and precise assessment of LTT caliber changes. Minor LTT caliber changes can be observed even in normals and, in case of an otherwise normal S-CT study, can be regarded as artifacts.

A new algorithm for blind restoration of space-invariant image degradations is presented. Once the singular value decomposition (SVD) of the degraded image is computed, degradation system identification and image restoration are performed through operations on singular vectors and singular values of the degraded image. The unknown point-spread function (PSF) is estimated from the singular vectors, while noise variance is estimated from the smallest singular values. It has been shown how to estimate the phase function for blurs that have zeros on the unit bicircle. Image restoration is performed through Wiener filtering of singular vectors. Knowledge about statistical properties of singular vectors and singular values is built into estimation and restoration procedures.

In this paper we propose a new blur identification algorithm based on singular value decomposition (SVD) of degraded images. An unknown space-invariant point-spread function (PSF) is also decomposed using SVD. Magnitude functions of PSF singular vectors (left and right) are identified using averaged spectra of corresponding singular vectors of the degraded image. Phase functions of PSF singular vectors are supposed to be zero, except for the case when zero crossings can be detected from corresponding magnitude functions. In the proposed method, the two dimensional PSF estimation procedure is decomposed into several one-dimensional estimation procedures. The PSF estimation algorithm does not require numerical optimization, suggesting a fast and straightforward procedure.

The paper presents a novel method for CT head image automatic segmentation. The images are obtained from patients having a spontaneous intra-cerebral brain hemorrhage (ICH). The results of the segmentation are images partitioned into five regions of interest corresponding to four tissue classes (skull, brain, calcifications and ICH) and background. Once the images are segmented it is possible to calculate various hemorrhage region parameters such as size, position, etc. The segmentation is performed in three major steps. In the first phase feature extraction and normalization is performed using a receptive field (RF). Experiments were performed to determine the optimal RF structure. Pixels are classified in the second phase using the radial basis function (RBF) artificial neural network. Experiments with different RBF network topologies were performed in order to determine the optimal basis functions, network size and a training algorithm. The segmentation results obtained using the RBF network were compared with results obtained by multi-layer perceptron neural network (MLP). In the third phase the image regions obtained by the RBF network were labeled using an expert system. Experiments have shown that the proposed method successfully performs image segmentation.

A new 3D technique for the human spontaneous intracerebral brain hemorrhage (ICH) region segmentation and quantification is presented in this paper. The ICH primary region segmentation algorithm uses the K-means histogram-based clustering algorithm. The ICH edema region segmentation algorithm employs an iterative morphological processing of the ICH brain data. A volume rendering technique is used for the effective 3D visualization of ICH segmented regions. A computer program is developed for use in the human spontaneous ICH study involving large number of patients. Some experimental measurements and visualization results are presented which were computed on real ICH patient brain data.>