Deep learning has proven to be one of the most effective methods in analyzing biological images to extract parameters fundamental for studying physiological functions and pathological conditions. In particular, when coupled with time-lapse microscopy (TLM), deep learning proves particularly effective in studying behaviors involving temporal dynamics. However, TLM videos are often affected by experimental noise and setup limitations, which can lead to inaccurate and poorly reproducible results. Taking advantage of the variational and generative capabilities of Variational Autoencoders (VAEs), we propose VAE-MOTION, a deep learning-based model for the analysis of cardiac contractile dynamics. By incorporating a temporal encoder into its architecture, our model allows the restoration of video quality by removing noise or increasing resolution, while simultaneously extracting accurate contraction-related signals from the latent space. The generation of synthetic videos allowed extensive training of VAE-MOTION, which subsequently validated on real videos from two different cardiac tissue models: 2D monolayers and 3D microtissues. VAE-MOTION was compared to two gold-standard methods in extracting contraction parameters relevant to drug efficacy or toxicity studies, demonstrating its potential for analyzing temporal dynamics in a given phenomenon or process.

The Anti-Nuclear Antibodies (ANA) test using Human Epithelial type 2 (HEp-2) cells in the Indirect Immuno-Fluorescence (IIF) assay protocol is considered the gold standard for detecting Connective Tissue Diseases. Computer-assisted systems for HEp-2 image analysis represent a growing field that harnesses the potential offered by novel machine learning techniques to address the classification of HEp-2 images and ANA patterns. In this study, we introduce an innovative platform based on transfer learning with pre-trained deep learning models. This platform combines the power of unsupervised deep description of HEp-2 images, a novel feature selection approach designed for unbalanced datasets, and independent testing using two distinct datasets from different hospitals to tackle cross-hardware compatibility issues. To enhance the trustworthiness of our method, we also present a modified version of gradient-weighted class activation mapping for regional explainability and introduce a new sample quality index based on the Jensen-Shannon divergence to enhance method reliability and quantify sample heterogeneity. The results we provide demonstrate exceptionally high performance in intensity and ANA pattern recognition when compared to state-of-the-art approaches. Our method's ability to eliminate the need for cell segmentation in favor of statistical analysis of the sample makes it applicable, robust, and versatile. Our future work will focus on addressing the challenge of mitotic spindle recognition by expanding our proposed approach to cover mixed patterns. • A novel AI platform for the classification of HEp-2 specimens without cell segmentation • Loss-based Multiclass Area Under the Curve feature selection approach • A novel Grad-CAM explainability method for transfer learning using pre-trained DL architectures • Novel sample quality index based on the Jensen-Shannon divergence index

MOTIVATION: In fluorescence microscopy, single-molecule localization microscopy (SMLM) techniques aim at localizing with high-precision high-density fluorescent molecules by stochastically activating and imaging small subsets of blinking emitters. Super resolution plays an important role in this field since it allows to go beyond the intrinsic light diffraction limit. RESULTS: In this work, we propose a deep learning-based algorithm for precise molecule localization of high-density frames acquired by SMLM techniques whose ℓ2-based loss function is regularized by non-negative and ℓ0-based constraints. The ℓ0 is relaxed through its continuous exact ℓ0 (CEL0) counterpart. The arising approach, named DeepCEL0, is parameter-free, more flexible, faster and provides more precise molecule localization maps if compared to the other state-of-the-art methods. We validate our approach on both simulated and real fluorescence microscopy data. AVAILABILITY AND IMPLEMENTATION: DeepCEL0 code is freely accessible at https://github.com/sedaboni/DeepCEL0.

Abstract In this paper, we present the novel Deep-MEG approach in which image-based representations of magnetoencephalography (MEG) data are combined with ensemble classifiers based on deep convolutional neural networks. For the scope of predicting the early signs of Alzheimer’s disease (AD), functional connectivity (FC) measures between the brain bio-magnetic signals originated from spatially separated brain regions are used as MEG data representations for the analysis. After stacking the FC indicators relative to different frequency bands into multiple images, a deep transfer learning model is used to extract different sets of deep features and to derive improved classification ensembles. The proposed Deep-MEG architectures were tested on a set of resting-state MEG recordings and their corresponding magnetic resonance imaging scans, from a longitudinal study involving 87 subjects. Accuracy values of 89% and 87% were obtained, respectively, for the early prediction of AD conversion in a sample of 54 mild cognitive impairment subjects and in a sample of 87 subjects, including 33 healthy controls. These results indicate that the proposed Deep-MEG approach is a powerful tool for detecting early alterations in the spectral–temporal connectivity profiles and in their spatial relationships.

The emerging tumor-on-chip (ToC) approaches allow to address biomedical questions out of reach with classical cell culture techniques: in biomimetic 3D hydrogels they partially reconstitute ex vivo the complexity of the tumor microenvironment and the cellular dynamics involving multiple cell types (cancer cells, immune cells, fibroblasts, etc.). However, a clear bottleneck is the extraction and interpretation of the rich biological information contained, sometime hidden, in the cell co-culture videos. In this work, we develop and apply novel video analysis algorithms to automatically measure the cytotoxic effects on human cancer cells (lung and breast) induced either by doxorubicin chemotherapy drug or by autologous tumor-infiltrating cytotoxic T lymphocytes (CTL). A live fluorescent dye (red) is used to selectively pre-stain the cancer cells before co-cultures and a live fluorescent reporter for caspase activity (green) is used to monitor apoptotic cell death. The here described open-source computational method, named STAMP (spatiotemporal apoptosis mapper), extracts the temporal kinetics and the spatial maps of cancer death, by localizing and tracking cancer cells in the red channel, and by counting the red to green transition signals, over 2-3 days. The robustness and versatility of the method is demonstrated by its application to different cell models and co-culture combinations. Noteworthy, this approach reveals the strong contribution of primary cancer-associated fibroblasts (CAFs) to breast cancer chemo-resistance, proving to be a powerful strategy to investigate intercellular cross-talks and drug resistance mechanisms. Moreover, we defined a new parameter, the 'potential of death induction', which is computed in time and in space to quantify the impact of dying cells on neighbor cells. We found that, contrary to natural death, cancer death induced by chemotherapy or by CTL is transmissible, in that it promotes the death of nearby cancer cells, suggesting the release of diffusible factors which amplify the initial cytotoxic stimulus.

Research on automatic emotion recognition from speech has recently focused on the prediction of time-continuous dimensions (e.g., arousal and valence) of spontaneous and realistic expressions of emotion, as found in real-life interactions. However, the automatic prediction of such emotions poses several challenges, such as the subjectivity found in the definition of a gold-standard from a pool of raters and the issue of data scarcity in training models. In this work, we introduce a novel emotion recognition system, based on ensembles of single-speaker-regression-models. The estimation of emotion is provided by combining a subset of the initial pool of single-speaker-regression-models selecting those that are most concordant among them. The proposed approach allows the addition or removal of speakers from the ensemble without the necessity to re-build the entire recognition system. The simplicity of this aggregation strategy, coupled with the flexibility assured by the modular architecture, and the promising results observed on the RECOLA database highlight the potential implications of the proposed method in a real-life scenario and in particular in web-based applications.

We hypothesize that quantification of structural similarity or dissimilarity between paired mammographic regions can be effective in detecting asymmetric signs of breast cancer. Bilateral masking procedures are applied for this purpose by using automatically detected anatomical landmarks. Changes in structural information of the extracted regions are investigated using spherical semivariogram descriptors and correlation-based structural similarity indices in the spatial and complex wavelet domains. The spatial distribution of grayscale values as well as of the magnitude and phase responses of multidirectional Gabor filters are used to represent the structure of mammographic density and of the directional components of breast tissue patterns, respectively. A total of 188 mammograms from the DDSM and mini-MIAS databases, consisting of 47 asymmetric cases and 47 normal cases, were analyzed. For the combined dataset of mammograms, areas under the receiver operating characteristic curves of 0.83, 0.77, and 0.87 were obtained, respectively, with linear discriminant analysis, the Bayesian classifier, and an artificial neural network with radial basis functions, using the features selected by stepwise logistic regression and leave-one-patient-out cross-validation. Two-view analysis provided accuracy up to 0.94, with sensitivity and specificity of 1 and 0.88, respectively.

In this work we propose a joint two-side masking procedure for automatic analysis of mammographic images. The primary objective is the improvement of computerized systems capability in revealing additional findings, as the asymmetrical changes of the breast parenchyma. The method allows the proper comparison of the left and right breast by progressive selection of paired small areas on the mammogram, primarily the so-called ''forbidden areas", zones that need special attention in mammographic interpretation. The masking of specific areas of the mammogram requires the identification of two anatomical structures of the breast: the pectoral muscle and the nipple used, together with the breast skin line, to find paired matching points on the images for comparison. With this purpose, specific algorithms have been developed. In particular, a new method for nipple extraction will be presented and validated by expert radiologists, by the use of a proprietary program developed by the authors. Finally, an application example of the automatic Tabar masking procedure will be shown, in order to point out the potential of this method in detection of suspicious areas in mammograms.

Uncertainty modelling in measurement represents a crucial task since the final result of a measurement process cannot be expressed by a single value, but by a distribution of values over an interval within which the measurements lie with a given confidence level. A classical Type-1 fuzzy set could be the natural choice for uncertainty model, since a Membership Function (MF) intrinsically embeds the concepts of confidence interval and confidence level. Moreover, from computational aspects, working on fuzzy sets is much more easily than a Montecarlo simulation, that is actually the recommended approach. However, both the approaches need reliable and specific assumptions on the probability distribution of the input variables. So, in many practical cases, a Type-2 fuzzy set is needed in order to overcome this problem. In this paper, we will provide a full description of how a Type-2 MF can be built in order to model the uncertainty of a variable and how to evaluate uncertainty propagation through a generic function. A practical example of this representation will be also provided and compared with a Montecarlo simulation.

Mammography is the most effective method for early detection of breast diseases. However, the typical diagnostic signs, such as masses and microcalcifications, are difficult to be detected because mammograms are low contrast and noisy images. In this paper, we present an algorithm for mammographic images enhancement and denoising based on the wavelet transform. In particular, we develop an adaptive procedure to perform an optimal denoising using a local iterative fuzzy noise variance estimation. Moreover, the degree of enhancement is adaptively tuned at each scale. The proposed algorithm has been tested on phantom and clinical images.

A fuzzy logic system (FLS) can be completely described through its behavioral function. If the analytical expression of this function is known, then the system synthesis consists in the realization of a structure implementing it in the most convenient way. So, the problem of the FLS design can be considered as a function approximation problem. Many methods have been proposed in literature to minimize a certain error parameter. Nevertheless, sometimes the specifications of the problem need some further analytical properties of the approximating function. This paper analyzes multiple-input-single-output (MISO) FLSs with polynomial membership functions inspecting the case of two input variables. In particular, it will be shown what characteristics such systems should have in order to respect the constraints imposed on the derivatives of the implemented function, such as their continuity up to a desired order and even their value in the grid points of the input space. Furthermore, we prove that these systems can uniformly approximate, under certain conditions, any multivariate function with a desired approximation accuracy both on the target function and its partial derivatives. Proper illustrative examples will show the behavior of these FLSs comparing their performance with other methods proposed in literature.

Noise estimation is an important issue in image processing because it is a fundamental step in many algorithms for noise suppression and then for image restoration. In the literature, many approaches have been presented in order to obtain good results. This paper presents a novel method suitable for obtaining a good estimation if the type of noise distribution is known. In particular, the algorithm provides the variance of the noise distribution and the proof that the distribution itself matches the foreseen one. The algorithm has been tested on different images affected by Gaussian noise. The simulations show results better than those obtained with other approaches.