Pre-designed Features Research Articles

IIFDD: Intra and inter-modal fusion for depression detection with multi-modal information from Internet of Medical Things

Depression is now a prevalent mental illness and multimodal data-based depression detection is an essential topic of research. Internet of Medical Things devices can provide data resources such as text, audio, and vision, which is valuable for depression detection. Moreover, previous studies have concentrated on using single characteristics of each modality, such as low-dimensional pre-designed features and high-level deep representation, which cannot completely capture the emotional information included in the data. Against this background, we design an intra-modal and inter-modal fusion framework called IIFDD for Corpus-based depression detection. Intra-modal fusion module is designed to integrate low-dimensional pre-designed features and high-dimension deep representation from the same modality for better learning of the semantics information. Then, the inter-modal fusion module is proposed to fuse features from different modalities with attention mechanisms and use the fused result to complete the depression classification. Experiments on two Chinese depression corpus datasets with acoustics, textual, and visual features show that IIFDD can achieve state-of-the-art performance for depression detection.

Amazon wood species classification: a comparison between deep learning and pre-designed features

In many countries, the wood industry is a crucial sector and has a significant economic impact. In this sense, illegal logging is a way to reduce costs, avoiding taxes, or having access to more valuable wood species. To combat the latter, the recognition of wood species is crucial. However, this task is usually performed by experts through visual inspection, a process that requires sanding and cleaning the wood surface, and an impractical task for use in the field. In this paper, the acquisition process was simplified and a new wood dataset was introduced, where a simple pocket knife cut is used to expose the timber section for inspection. Four deep learning models with transfer learning were investigated and compared with traditional pre-designed feature methods. Additionally, the models were evaluated with a cross-validation scheme to avoid any bias. The experimental results show that deep learning outperforms pre-design features for wood classification. DenseNet achieved 98.13% of accuracy, indicating that it could be applied to assist untrained agents in wood classification.

An analysis of timber sections and deep learning for wood species classification

The wood species classification is an essential field of investigation that can help to combat illegal logging, then providing the timber certification and allowing the application of correct timber taxing. Today, the wood classification relies on highly qualified professionals that analyze texture patterns on timber sections. However, these professionals are scarce, costly, and subject to failure. Therefore, the automation of this task using computational methods is promising. Deep learning has proven to be the ultimate technique in computer vision tasks, but it has not been much exploited to perform timber classification due to the difficulty of building large databases to train such networks. In this study, we introduced the biggest data set of wood timber microscope images to the date, with 281 species, having three types of timber sections: transverse, radial, and tangential. We investigated the use of transfer learning from pre-trained deep neural networks for wood species classification and compared their results with a state-of-art pre-designed feature method. The experimental results show that traverse section images using a densely connected network achieved 98.7% of correct classification against 85.9% of standard pre-designed features.

ML–ResNet: A novel network to detect and locate myocardial infarction using 12 leads ECG

Background and objectiveMyocardial infarction (MI) is one of the most threatening cardiovascular diseases for human beings, which can be diagnosed by electrocardiogram (ECG). Automated detection methods based on ECG focus on extracting handcrafted features. However, limited by the performance of traditional methods and individual differences between patients, it's difficult for predesigned features to detect MI with high performance. MethodsThe paper presents a novel method to detect and locate MI combining a multi-lead residual neural network (ML–ResNet) structure with three residual blocks and feature fusion via 12 leads ECG records. Specifically, single lead feature branch network is trained to automatically learn the representative features of different levels between different layers, which exploits local characteristics of ECG to characterize the spatial information representation. Then all the lead features are fused together as global features. To evaluate the generalization of proposed method and clinical utility, two schemes including the intra-patient scheme and inter-patient scheme are all employed. ResultsExperimental results based on PTB (Physikalisch-Technische Bundesanstalt) database shows that our model achieves superior results with the accuracy of 95.49%, the sensitivity of 94.85%, the specificity of 97.37%, and the F1 score of 96.92% for MI detection under the inter-patient scheme compared to the state-of-the-art. By contrast, the accuracy is 99.92% and the F1 score is 99.94% based on 5-fold cross validation under the intra-patient scheme. As for five types of MI location, the proposed method also yields an average accuracy of 99.72% and F1 of 99.67% in the intra-patient scheme. ConclusionsThe proposed method for MI detection and location has achieved superior results compared to other detection methods. However, further promotion of the performance based on MI location for the inter-patient scheme still depends significantly on the mass data and the novel model which reflects spatial location information of different leads subtly.

From Handcrafted to Deep-Learning-Based Cancer Radiomics: Challenges and Opportunities

Recent advancements in signal processing and machine learning coupled with developments of electronic medical record keeping in hospitals and the availability of extensive set of medical images through internal/external communication systems, have resulted in a recent surge of significant interest in "Radiomics". Radiomics is an emerging and relatively new research field, which refers to extracting semi-quantitative and/or quantitative features from medical images with the goal of developing predictive and/or prognostic models, and is expected to become a critical component for integration of image-derived information for personalized treatment in the near future. The conventional Radiomics workflow is typically based on extracting pre-designed features (also referred to as hand-crafted or engineered features) from a segmented region of interest. Nevertheless, recent advancements in deep learning have caused trends towards deep learning-based Radiomics (also referred to as discovery Radiomics). Considering the advantages of these two approaches, there are also hybrid solutions developed to exploit the potentials of multiple data sources. Considering the variety of approaches to Radiomics, further improvements require a comprehensive and integrated sketch, which is the goal of this article. This manuscript provides a unique interdisciplinary perspective on Radiomics by discussing state-of-the-art signal processing solutions in the context of Radiomics.

Deep Learning-Based Proarrhythmia Analysis Using Field Potentials Recorded From Human Pluripotent Stem Cells Derived Cardiomyocytes

An early characterization of drug-induced cardiotoxicity may be possible by combining comprehensive in vitro proarrhythmia assay and deep learning techniques. We aimed to develop a method to automatically detect irregular beating rhythm of field potentials recorded from human pluripotent stem cells (hPSC) derived cardiomyocytes (hPSC-CM) by multi-electrode array (MEA) system. We included field potentials from 380 experiments, which were labeled as normal or arrhythmic by electrophysiology experts. Convolutional and recurrent neural networks (CNN and RNN) were employed for automatic classification of field potential recordings. A preparation phase was initially applied to split 60-s long recordings into a series of 5-s windows. Subsequently, the classification phase comprising of two main steps was designed and applied. The first step included the classification of 5-s windows by using a designated CNN. While, the results of 5-s window assessments were used as the input sequence to an RNN that aggregates these results in the second step. The output was then compared to electrophysiologist-level arrhythmia detection, resulting in 0.83 accuracy, 0.93 sensitivity, 0.70 specificity, and 0.80 precision. In summary, this paper introduces a novel method for automated analysis of “irregularity” in an in vitro model of cardiotoxicity experiments. Thus, our method may overcome the drawbacks of using predesigned features that restricts the classification performance to the comprehensiveness and the quality of the designed features. Furthermore, automated analysis may facilitate the quality control experiments through the procedure of drug development with respect to cardiotoxicity and avoid late drug attrition from market.

Controllable silicon nano-grass formation using a hydrogenation assisted deep reactive ion etching

Highly controllable silicon nano-grass formation is reported based on a hydrogen assisted reactive ion etching method in desirable shapes and locations. By controlling the etching parameters, one can achieve grass-free high aspect ratio vertical or three-dimensional structures on silicon substrates. On the other hand, one can program the etching procedure to arrive at grass-full surfaces and structures in pre-designed features and in desired places. The improved wetting properties of the grass-full surfaces have been investigated. In addition, the grass-full surface has been used to entrap kidney cells. Aspect ratios of the order of 40–50 and features of the size of 250 nm can be achieved.

Formation of silicon nanograss and microstructures on silicon using deep reactive ion etching

A hydrogen-assisted reactive ion etching method for the silicon nanograss formation in desirable shapes and locations is reported. By controlling the etching parameters of this process, one can achieve grass-free high aspect ratio vertical structures on silicon substrates or arrive at grass-full structures in pre-designed features and in desired places. Using this method vertical or spaghetti-like Si nanograss with a height of 2–3 µm and width of 30–100 nm is realised. The effect of SF6 as a critical parameter on the nanograss formation has been studied relating the grass density and width to pressure and power. Inclusion of grass onto the silicon surface changes its wetting parameter to make it hydrophilic and oleophilic. The resonance frequency of silicon membranes with the addition of nanograss has shown a significant drop because of an improved surface adhesion to oil, owing to increasing the effective mass of the membrane.

Fabrication of microsensors using unmodified office inkjet printers

In this paper, a simple microfabrication technique for delivering biological macromolecules and patterning microelectrode arrays using desktop inkjet printers is described. In particular, we have demonstrated that: (1) phospholipid films can be patterned into pre-designed features (dots, lines and grids) with a resolution down to 50 μm; (2) miniaturized electrode sets can be prepared via “printing” an alkanethiolate self-assembled monolayer on a gold CD-R and subsequent wet chemical etching; (3) multiplex chemical sensors (e.g., for iodine vapor) can be fabricated using a multi-step inkjet printing, selective etching and polymer deposition protocol. It is remarkable that the entire fabrication process has been carried out in a regular wet laboratory setting, for which no hardware or software modifications to the commercial printer are required and regular CD-Rs/glass slides are used as substrate materials. This technique outlines a generic, “amateur”-friendly approach for the rapid and economical construction of disposable multiplex microsensors and portable electroanalytical units.