ABSTRACT Recent trends in computational pathology favour increasingly complex deep learning architectures, raising the question of whether such complexity is necessary for routine diagnostic tasks. This study challenges this assumption through a comprehensive analysis of the relationship between model complexity, data pre‐processing, and performance across four fundamental digital pathology tasks: nuclei counting, steatosis quantification, glomeruli detection, and Ki67 proliferation index (PI) assessment. We evaluated five deep learning models of varying complexity (lightweight: MobileNetV2, U‐Net, and more complex: ConvNeXt, K‐Net, and Swin Transformer) combined with different image pre‐processing techniques. To evaluate model performance without extensive ground truth (GT) annotations, we introduced a validation strategy utilizing the relative absolute deviation (RAD) between network predictions and correlation of performance metrics. Our findings demonstrate that pre‐processing strategies, particularly stain normalization (NORM), can be more impactful than model complexity, reducing error rates by up to 50% compared to processing original (ORIG) images. With appropriate pre‐processing, lightweight models achieved comparable or superior results to complex models while reducing processing times by up to 40%. Only specific tasks involving complex morphological features, such as glomeruli detection, significantly benefited from more sophisticated architectures. This study provides an evidence‐based framework for selecting optimal model‐pre‐processing combinations in clinical settings, suggesting that investing in pre‐processing pipelines rather than model complexity may be more beneficial for routine computational pathology applications.

Time series forecasting remains challenging in the presence of nonstationarity, regime changes, and observation noise. Many existing machine learning approaches rely on complex architectures that often lead to unstable training and limited robustness. To address these limitations, we propose QAAR-SIREN, a compact forecasting framework that improves stability through residual learning and complementary feature representations. Instead of predicting absolute values, the model forecasts temporal increments, mitigating nonstationarity effects. It integrates three information sources: raw temporal lags, attention-based contextual summarization, and lightweight nonlinear features extracted from a shallow variational quantum circuit applied to the most recent observation. The quantum component functions as a compact nonlinear feature extractor that enriches the input representation without increasing architectural complexity. Experiments on synthetic signals with regime transitions and heterogeneous noise, as well as real-world datasets from climate, energy demand, finance, and transportation, demonstrate that QAAR-SIREN achieves strong and stable predictive performance. The model attains coefficients of determination up to approximately 0.985 with low mean squared error. Ablation studies confirm that observed gains arise from the complementary effects of residual learning, attention-based context aggregation, and quantum feature extraction.

The automatic design of a 3D tooth model plays a crucial role in dental digitization. However, current approaches face challenges in compositional 3D tooth generation because both the layouts and shapes of missing teeth need to be optimized. In addition, collision conflicts are often omitted in 3D Gaussian-based compositional 3D generation, where objects may intersect with each other due to the absence of explicit geometric information on the object surfaces. Motivated by graph generation through diffusion models and collision detection using 3D Gaussians, we propose an approach named DM-CFO for compositional tooth generation, where the layout of missing teeth is progressively restored during the denoising phase under both text and graph constraints. Then, the Gaussian parameters of each layout-guided tooth and the entire jaw are alternately updated using score distillation sampling (SDS). Furthermore, a regularization term based on the distances between the 3D Gaussians of neighboring teeth and the anchor tooth is introduced to penalize tooth intersections. Experimental results on three tooth-design datasets demonstrate that our approach significantly improves the multiview consistency and realism of the generated teeth compared with existing methods.

Malicious firmware upgrading represents a critical security vulnerability in Internet of Things (IoT) devices. This study introduces HyCNNAt, a novel hybrid deep learning network for IoT malware detection that synergistically combines Convolutional Neural Networks (CNNs) with transformer attention mechanisms. HyCNNAt’s architecture vertically and horizontally stacks convolution and attention layers, enhancing the network’s generalization capabilities, capacity, and overall effectiveness. We evaluated HyCNNAt using a publicly available IoT firmware dataset, where it demonstrated superior performance with the highest accuracy ([Formula: see text]), F1-score ([Formula: see text]), and recall ([Formula: see text]), highlighting its robust classification capabilities, although its precision ([Formula: see text]) exhibited variability compared to state-of-the-art models such as CoAtNet, MobileViT, MobileNet, and MobileNet variants using transfer learning. These results underscore HyCNNAt’s potential as a robust solution for addressing the pressing challenge of IoT malware detection.

Abstract Grapevine leaf is a commodity that is collected only once a year and has a high return on investment due to its export. However, only certain types of grapevine leaves are consumed. Therefore, it is extremely important to distinguish the types of grapevine leaves. In particular, performing this process automatically on industrial machines will reduce human errors, workload, and thus cost. In this study, a new hybrid approach based on a convolutional neural network is proposed that can automatically distinguish the types of grapevine leaves. In the proposed approach, firstly, the overfitting of network models is prevented by applying data augmentation techniques. Second, new synthetic images were created with the ESRGAN technique to obtain detailed texture information. Third, the top blocks of the MobileNetV2 and VGG19 CNN models were replaced with the newly designed top block, effectively extracting features with the data. Fourthly, the GASVM algorithm was adapted and used to create a subset of the features to eliminate the ineffective and unimportant ones from the obtained features. Finally, SVM classification was performed with the feature subset consisting of 314 features, and approximately 2% higher accuracy and MCC score were obtained compared to the approaches in the literature.

The use of Convolutional Neural Networks (CNN) pre-trained for a particular task, as a feature extractor for an alternate task, is a standard practice in many image classification paradigms. However, to date there have been comparatively few works exploring this technique for speech classification tasks. Herein, we utilise a pre-trained end-to-end Automatic Speech Recognition CNN as a feature extractor for the task of food-type recognition from speech. Furthermore, we also explore the benefits of Compact Bilinear Pooling for combining multiple feature representations extracted from the CNN. Key results presented indicate the suitability of this approach. When combined with a Recurrent Neural Network classifier, our strongest system achieves, for a seven-class food-type classification task an unweighted average recall of 73.3% on the test set of the iHEARu-EAT database.

A novel method to detect human falls in depth videos is presented in this paper. A fast and robust shape sequence descriptor, namely the Silhouette Orientation Volume (SOV), is used to represent actions and classify falls. The SOV descriptor provides high classification accuracy even with a combination of simple associated models, such as Bag-of-Words and the Naïve Bayes classifier. Experiments on the public SDU-Fall dataset show that this new approach achieves up to 91.89% fall detection accuracy with a single-view depth camera. The classification rate is about 5% higher than the results reported in the literature. An overall accuracy of 89.63% was obtained for the six-class action recognition, which is about 25% higher than the state of the art. Moreover, a perfect silhouette-based action recognition rate of 100% is achieved on the Weizmann action dataset.

In this letter, a two-stage method for airport detection on remote sensing images is proposed. In the first stage, a new algorithm composed of several line-based processing steps is used for extraction of candidate airport regions. In the second stage, the scale-invariant feature transformation and Fisher vector coding are used for efficient representation of the airport and nonairport regions and support vector machines employed for classification. In order to evaluate the performance of the proposed method, extensive experiments are conducted on airports around the world with different layouts. The measures used in the evaluation are accuracy, sensitivity, and specificity. The proposed method achieved an accuracy of 94.6%, which was benchmarked with two previous methods to prove its superiority.