Convolutional neural networks (CNNs) have gained prominence in the research literature on image classification over the last decade. One shortcoming of CNNs, however, is their lack of generalizability and tendency to overfit when presented with small training sets. Augmentation directly confronts this problem by generating new data points providing additional information. In this paper, we investigate the performance of more than ten different sets of data augmentation methods, with two novel approaches proposed here: one based on the discrete wavelet transform and the other on the constant-Q Gabor transform. Pretrained ResNet50 networks are finetuned on each augmentation method. Combinations of these networks are evaluated and compared across four benchmark data sets of images representing diverse problems and collected by instruments that capture information at different scales: a virus data set, a bark data set, a portrait dataset, and a LIGO glitches data set. Experiments demonstrate the superiority of this approach. The best ensemble proposed in this work achieves state-of-the-art (or comparable) performance across all four data sets. This result shows that varying data augmentation is a feasible way for building an ensemble of classifiers for image classification.

Research in sound classification and recognition is rapidly advancing in the field of pattern recognition. One important area in this field is environmental sound recognition, whether it concerns the identification of endangered species in different habitats or the type of interfering noise in urban environments. Since environmental audio datasets are often limited in size, a robust model able to perform well across different datasets is of strong research interest. In this paper, ensembles of classifiers are combined that exploit six data augmentation techniques and four signal representations for retraining five pre-trained convolutional neural networks (CNNs); these ensembles are tested on three freely available environmental audio benchmark datasets: (i) bird calls, (ii) cat sounds, and (iii) the Environmental Sound Classification (ESC-50) database for identifying sources of noise in environments. To the best of our knowledge, this is the most extensive study investigating ensembles of CNNs for audio classification. The best-performing ensembles are compared and shown to either outperform or perform comparatively to the best methods reported in the literature on these datasets, including on the challenging ESC-50 dataset. We obtained a 97% accuracy on the bird dataset, 90.51% on the cat dataset, and 88.65% on ESC-50 using different approaches. In addition, the same ensemble model trained on the three datasets managed to reach the same results on the bird and cat datasets while losing only 0.1% on ESC-50. Thus, we have managed to create an off-the-shelf ensemble that can be trained on different datasets and reach performances competitive with the state of the art.

High-throughput in vitro drug assays have been impacted by recent advances in human induced pluripotent stem cell-derived cardiomyocyte (hiPSC-CM) technology and by contact-free all-optical systems simultaneously measuring action potentials (APs) and Ca²⁺ transients (CaTrs). Parallel computational advances have shown that in silico simulations can predict drug effects with high accuracy. We combine these in vitro and in silico technologies and demonstrate the utility of high-throughput experimental data to refine in silico hiPSC-CM populations and to predict and explain drug action mechanisms. Optically obtained hiPSC-CM APs and CaTrs were used from spontaneous activity and under optical pacing in control and drug conditions at multiple doses. An updated version of the Paci2018 model was developed to refine the description of hiPSC-CM spontaneous electrical activity; a population of in silico hiPSC-CMs was constructed and calibrated using simultaneously recorded APs and CaTrs. We tested in silico five drugs (astemizole, dofetilide, ibutilide, bepridil, and diltiazem) and compared the outcomes to in vitro optical recordings. Our simulations showed that physiologically accurate population of models can be obtained by integrating AP and CaTr control records. Thus, constructed population of models correctly predicted the drug effects and occurrence of adverse episodes, even though the population was optimized only based on control data and in vitro drug testing data were not deployed during its calibration. Furthermore, the in silico investigation yielded mechanistic insights; e.g., through simulations, bepridil's more proarrhythmic action in adult cardiomyocytes compared to hiPSC-CMs could be traced to the different expression of ion currents in the two. Therefore, our work 1) supports the utility of all-optical electrophysiology in providing high-content data to refine experimentally calibrated populations of in silico hiPSC-CMs, 2) offers insights into certain limitations when translating results obtained in hiPSC-CMs to humans, and 3) shows the strength of combining high-throughput in vitro and population in silico approaches.

The growing importance of human induced pluripotent stem cell-derived cardiomyoyctes (hiPSC-CMs), as patient-specific and disease-specific models for studying cellular cardiac electrophysiology or for preliminary cardiotoxicity tests, generated better understanding of hiPSC-CM biophysical mechanisms and great amount of action potential and calcium transient data. In this paper, we propose a new hiPSC-CM in silico model, with particular attention to Ca²⁺ handling. We used (i) the hiPSC-CM Paci2013 model as starting point, (ii) a new dataset of Ca²⁺ transient measurements to tune the parameters of the inward and outward Ca²⁺ fluxes of sarcoplasmic reticulum, and (iii) an automatic parameter optimization to fit action potentials and Ca²⁺ transients. The Paci2018 model simulates, together with the typical hiPSC-CM spontaneous action potentials, more refined Ca²⁺ transients and delayed afterdepolarizations-like abnormalities, which the old Paci2013 was not able to predict due to its mathematical formulation. The Paci2018 model was validated against (i) the same current blocking experiments used to validate the Paci2013 model, and (ii) recently published data about effects of different extracellular ionic concentrations. In conclusion, we present a new and more versatile in silico model, which will provide a platform for modeling the effects of drugs or mutations that affect Ca²⁺ handling in hiPSC-CMs.

In recent years, binary coding of image features, such as local binary patterns and local phase quantization, have become popular in a large variety of image quantification tasks. Lately, some non-binary codings, such as local ternary pattern, have been proposed to improve the performance of these binary based approaches. In these methods it is very important to correctly choose the thresholds applied for building the coding used to represent a given image and its features by a feature vector. In this work we compare several approaches for extracting local ternary/quinary pattern image features and ternary coding for local phase quantization on various types of biological microscope images using six image databases for sub-cellular and stem cell image classification. We use these image features for training a stand-alone support vector machine and a random subspace of support vector machines to separate the different classes present in each dataset. Moreover, several distance measures are tested. Our results show that, on the chosen datasets, the best approach uses a multi-threshold local quinary coding. The use of a more discriminating coding scheme than the binary one, combined with a pool of thresholds, helps in distinguishing descriptive features from noise, thus improving classification results. The Matlab code is available at bias.csr.unibo.it/nanni/TernaryCoding.rar.

Background and Purpose: Two new technologies are likely to revolutionize cardiac safety and drug development: in vitro experiments on human‐induced pluripotent stem cell‐derived cardiomyocytes (hiPSC‐CMs) and in silico human adult ventricular cardiomyocyte (hAdultV‐CM) models. Their combination was recently proposed as a potential replacement for the present hERG‐based QT study for pharmacological safety assessments. Here, we systematically compared in silico the effects of selective ionic current block on hiPSC‐CM and hAdultV‐CM action potentials (APs), to identify similarities/differences and to illustrate the potential of computational models as supportive tools for evaluating new in vitro technologies.Experimental Approach: In silico AP models of ventricular‐like and atrial‐like hiPSC‐CMs and hAdultV‐CM were used to simulate the main effects of four degrees of block of the main cardiac transmembrane currents.Key Results: Qualitatively, hiPSC‐CM and hAdultV‐CM APs showed similar responses to current block, consistent with results from experiments. However, quantitatively, hiPSC‐CMs were more sensitive to block of (i) L‐type Ca²⁺ currents due to the overexpression of the Na⁺/Ca²⁺ exchanger (leading to shorter APs) and (ii) the inward rectifier K⁺ current due to reduced repolarization reserve (inducing diastolic potential depolarization and repolarization failure).Conclusions and Implications: In silico hiPSC‐CMs and hAdultV‐CMs exhibit a similar response to selective current blocks. However, overall hiPSC‐CMs show greater sensitivity to block, which may facilitate in vitro identification of drug‐induced effects. Extrapolation of drug effects from hiPSC‐CM to hAdultV‐CM and pro‐arrhythmic risk assessment can be facilitated by in silico predictions using biophysically‐based computational models.