Classification Performance Research Articles

A simplified adversarial architecture for cross-subject silent speech recognition using electromyography

Objective. The decline in the performance of electromyography (EMG)-based silent speech recognition is widely attributed to disparities in speech patterns, articulation habits, and individual physiology among speakers. Feature alignment by learning a discriminative network that resolves domain offsets across speakers is an effective method to address this problem. The prevailing adversarial network with a branching discriminator specializing in domain discrimination renders insufficiently direct contribution to categorical predictions of the classifier.Approach. To this end, we propose a simplified discrepancy-based adversarial network with a streamlined end-to-end structure for EMG-based cross-subject silent speech recognition. Highly aligned features across subjects are obtained by introducing a Nuclear-norm Wasserstein discrepancy metric on the back end of the classification network, which could be utilized for both classification and domain discrimination. Given the low-level and implicitly noisy nature of myoelectric signals, we devise a cascaded adaptive rectification network as the front-end feature extraction network, adaptively reshaping the intermediate feature map with automatically learnable channel-wise thresholds. The resulting features effectively filter out domain-specific information between subjects while retaining domain-invariant features critical for cross-subject recognition.Main results. A series of sentence-level classification experiments with 100 Chinese sentences demonstrate the efficacy of our method, achieving an average accuracy of 89.46% tested on 40 new subjects by training with data from 60 subjects. Especially, our method achieves a remarkable 10.07% improvement compared to the state-of-the-art model when tested on 10 new subjects with 20 subjects employed for training, surpassing its result even with three times training subjects.Significance. Our study demonstrates an improved classification performance of the proposed adversarial architecture using cross-subject myoelectric signals, providing a promising prospect for EMG-based speech interactive application.

Med-MGF: multi-level graph-based framework for handling medical data imbalance and representation

BackgroundModeling patient data, particularly electronic health records (EHR), is one of the major focuses of machine learning studies in healthcare, as these records provide clinicians with valuable information that can potentially assist them in disease diagnosis and decision-making.MethodsIn this study, we present a multi-level graph-based framework called MedMGF, which models both patient medical profiles extracted from EHR data and their relationship network of health profiles in a single architecture. The medical profiles consist of several layers of data embedding derived from interval records obtained during hospitalization, and the patient-patient network is created by measuring the similarities between these profiles. We also propose a modification to the Focal Loss (FL) function to improve classification performance in imbalanced datasets without the need to imputate the data. MedMGF’s performance was evaluated against several Graphical Convolutional Network (GCN) baseline models implemented with Binary Cross Entropy (BCE), FL, class balancing parameter α\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\alpha$$\\end{document}, and Synthetic Minority Oversampling Technique (SMOTE).ResultsOur proposed framework achieved high classification performance (AUC: 0.8098, ACC: 0.7503, SEN: 0.8750, SPE: 0.7445, NPV: 0.9923, PPV: 0.1367) on an extreme imbalanced pediatric sepsis dataset (n=3,014, imbalance ratio of 0.047). It yielded a classification improvement of 3.81% for AUC, 15% for SEN compared to the baseline GCN+α\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\alpha$$\\end{document}FL (AUC: 0.7717, ACC: 0.8144, SEN: 0.7250, SPE: 0.8185, PPV: 0.1559, NPV: 0.9847), and an improvement of 5.88% in AUC and 22.5% compared to GCN+FL+SMOTE (AUC: 0.7510, ACC: 0.8431, SEN: 0.6500, SPE: 0.8520, PPV: 0.1688, NPV: 0.9814). It also showed a classification improvement of 3.86% for AUC, 15% for SEN compared to the baseline GCN+α\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\alpha$$\\end{document}BCE (AUC: 0.7712, ACC: 0.8133, SEN: 0.7250, SPE: 0.8173, PPV: 0.1551, NPV: 0.9847), and an improvement of 14.33% in AUC and 27.5% in comparison to GCN+BCE+SMOTE (AUC: 0.6665, ACC: 0.7271, SEN: 0.6000, SPE: 0.7329, PPV: 0.0941, NPV: 0.9754).ConclusionWhen compared to all baseline models, MedMGF achieved the highest SEN and AUC results, demonstrating the potential for several healthcare applications.

Elemental tracer determination and modeling for geographical origin designation of sweet oranges

Sweet oranges have long been an integral part of global health and culinary practices, offering a wealth of nutrients and bioactive compounds. Ensuring the authenticity of these citrus fruits is essential for maintaining consumer confidence, promoting transparency in sourcing, and protecting producers' reputations in the marketplace. In this study, we explored the feasibility of using multi-element profiling combined with pattern recognition algorithms to trace the origin of sweet orange samples. To achieve this, we employed an optimized microwave plasma atomic emission spectroscopy (MP-AES) method to analyze the elemental composition (Al, Ba, Ca, Cd, Co, Cr, Cu, Fe, K, Li, Mg, Mn, Mo, Na, Ni, Pb, Sr, and Zn) of 183 orange samples from four production regions in northeastern Argentina. Support vector machine (SVM), random forest (RF), and gradient boosting tree (GBT) models were then built using the collected data to identify elemental tracer's indicative of origin. Based on a comprehensive evaluation of overall accuracy, receiver operating characteristic (ROC) curves, and area under the curve (AUC), the GBT model demonstrated the best classification performance, achieving a 96.5 % correct prediction rate on test samples, as confirmed by the ROC curve (AUC = 0.973). Consequently, this approach provides compelling evidence for the potential utility of MP-AES combined with supervised modeling to determine the geographic origin of sweet oranges produced in Argentina, thereby contributing to consumer protection against fraud.

Robust mosquito species identification from diverse body and wing images using deep learning

Mosquito-borne diseases are a major global health threat. Traditional morphological or molecular methods for identifying mosquito species often require specialized expertise or expensive laboratory equipment. The use of convolutional neural networks (CNNs) to identify mosquito species based on images may offer a promising alternative, but their practical implementation often remains limited. This study explores the applicability of CNNs in classifying mosquito species. It compares the efficacy of body and wing depictions across three image collection methods: a smartphone, macro-lens attached to a smartphone and a professional stereomicroscope. The study included 796 specimens of four morphologically similar Aedes species, Aedes aegypti, Ae. albopictus, Ae. koreicus and Ae. japonicus japonicus. The findings of this study indicate that CNN models demonstrate superior performance in wing-based classification 87.6% (95% CI: 84.2–91.0) compared to body-based classification 78.9% (95% CI: 77.7–80.0). Nevertheless, there are notable limitations of CNNs as they perform reliably across multiple devices only when trained specifically on those devices, resulting in an average decline of mean accuracy by 14%, even with extensive image augmentation. Additionally, we also estimate the required training data volume for effective classification, noting a reduced requirement for wing-based classification compared to body-based methods. Our study underscores the viability of both body and wing classification methods for mosquito species identification while emphasizing the need to address practical constraints in developing accessible classification systems.Graphical abstract

Classification of painting styles based on the difference component

The digitisation and classification of artwork are crucial tasks in the management of art collections. However, the scarcity of some styles makes it difficult to obtain sufficient and representative samples in deep learning models, leading to sub-optimal classification performance. This study proposed a difference component for few-shot learning to improve the classification accuracy of existing models on few-shot data. First, the method pre-processes the sample data by permuting limited samples with sufficient data to generate related samples. Subsequently, sets of positive and negative samples with the same scale are generated, containing and not containing the limited samples. Second, a difference component model is constructed to learn the non-linear division among samples contained in the related samples. This is then combined with the existing classification model to determine whether the related samples contain limited samples. Finally, the types of each sample contained in related samples are analysed based on the types of related samples. The results of the multiple experiments demonstrate that the difference component can effectively learn the differential features between combinations of artworks. The combination of the existing classification model and the proposed difference component model has been demonstrated to effectively improve the classification accuracy of few-shot works. In addition, the classification result was superior to that of other baseline methods used for few-sample learning.

Optimized Tiny Machine Learning and Explainable AI for Trustable and Energy-Efficient Fog-Enabled Healthcare Decision Support System

The Internet of things (IoT)-based healthcare decision support system plays a crucial role in modern medicine, especially with the rise in chronic illnesses and an aging population necessitating continuous remote health monitoring. Current healthcare decision support systems struggle to deliver timely and accurate decisions with minimal latency due to limited real-time healthcare data and inefficient computational resources. There is a critical need for an optimized, energy-efficient machine learning model that reliably supports remote health monitoring within IoT and fog computing environments. Our study proposes an Optimized Tiny Machine Learning (TinyML) and Explainable AI (XAI) binary classification model for a trustable and energy-efficient healthcare decision support system, leveraging fog computing to optimize performance. The fog-based approach improves response times and enhances bandwidth usage, addressing critical needs such as reduced latency, higher bandwidth utilization, and decreased packet loss. To further improve efficiency, we incorporate the innovative mLZW data compression technique, significantly enhancing data communication efficiency and reducing response time to critical health alerts. However, limited real-time healthcare data records challenge machine learning classification performance. By implementing a TinyML algorithm, our system demonstrates superior performance to other machine learning models. The proposed optimized TinyML model achieves an impressive F1 score of 0.93 for health abnormalities detection, emphasizing its robustness and effectiveness. This paper highlights the potential of TinyML and XAI in delivering robust, trustworthy, and energy-aware healthcare solutions, making significant contributions toward effective remote health monitoring and decision support in fog-enabled IoT networks.Graphical abstract

Efficient Intrusion Detection in Patient Health Records Based on LSTM-Based RNN

Advancements in Information and Communication Technology applications play an important role in the field of healthcare and bio-medical engineering. The process of image processing and analysis starts from receiving visual information to giving out the description of the scene for improving the visual appearance of the human viewer and extract the features of the data. Moreover, IDS can be deployed along with other security mechanisms as a line of defence to ensure security of system and network resources. There have been various efforts in designing IDS using Machine Learning (ML) techniques. Additionally, by developing a hybrid strategy for intrusion detection and classification and using feature engineering approaches to extract significant features for learning, among other things, efforts have been made to improve the classification performance of ML-based IDS. However, as networking technology have advanced, attack types have evolved as well. For this reason, an efficient and successful intrusion detection and classification system must be created. To address the issue and achieve good generalization ability for intrusion detection and classification, the paper presents empirical analysis of LSTM based RNN classifiers. The suggested methods' effectiveness is assessed using a range of performance metrics, such as recall, f-score, accuracy, precision, and False Positive Rate (FPR). The LSTM based RNN method recognize attacks with 99.2% accuracy with minimum time complexity (5 s).

Divergence-guided disentanglement of view-common and view-unique representations for multi-view data

In the field of multi-view learning (MVL), it is crucial to extract both common (consistent) and unique (complementary) information across different views. While the focus has traditionally been on acquiring common information, there has been a recent shift towards exploring unique information as well. However, developing an MVL model that can simultaneously capture both common and unique information, thereby facilitating a comprehensive understanding of multi-view data, remains a significant challenge. To address this, we propose the Divergence-guided Multi-view Learning framework (DG-MVL), inspired by information-theoretic learning theory, specifically the generalized divergence measure. This framework employs multi-view autoencoders to disentangle the features obtained from each view into coarse common and unique components. By minimizing the divergence between the coarse common features learned from the common encoder of each view and maximizing the divergence between the coarse unique features from unique encoders simultaneously, we optimize the extraction of both common and unique information. Subsequently, we merge these features to generate a comprehensive and concise representation of the multi-view data, which can be easily utilized for various downstream tasks. We validate our framework on a synthetic multi-view dataset, demonstrating its effectiveness in disentangling common and unique information. Further experiments on various real-world datasets confirm the effectiveness of DG-MVL in capturing common and unique information from multi-view data, resulting in superior classification performance compared to existing methods. Code is available at https://github.com/LMFLRB/DG-MVL.git.

Predicting Severity Of Osteoarthritis Using Recurrent Convolutional Neural Networks (Rcnn) And Medical Imaging Data

Cartilage degeneration causing functional incapacity and discomfort defines common joint condition Osteoarthritis (OA). Correct classification of OA severity from knee X-ray images determines diagnosis and treatment course. Conventional methods of OA severity classification rely on hand inspection or basic machine learning techniques, which could not effectively capture the complex trends in imaging data. Using a dataset of 9,786 knee X-ray images, this work uses Recurrent Convolutional Neural Networks (RCNN) to project OA severity. Combining recurrent layers in the RCNN architecture helps to capture temporal correlations in spatial information, hence enhancing classification performance. Defining the dataset are five KL grades: 0 (healthy) to 4 (severe). This beats by around 5%, 7%, and 3%, respectively standard CNN, Deep Neural Network (DNN), and Deep Convolutional Neural Network (DCNN), with a precision of 87.2%, recall of 88.9%, accuracy of 89.5%, and F1-score of 88.0%.

Optimal scale combination selection in generalized multi-scale hybrid decision systems

When processing multi-scale hybrid data, the existing methods are insufficient to directly select the optimal scale for different types of data. To deal with this issue, the concept of generalized multi-scale hybrid decision systems (GMHDSs) consisting of numerical, nominal and set-valued values is first introduced. Then, under a scale combination, the information granules of different types of attributes in a GMHDS are constructed and they further fused as the granules of the system. Based on these granules of a given scale combination, positive region and conditional entropy are specified in a GMHDS. To select appropriate single scale systems to keep the positive region and the conditional entropy unchanged for final decision, concepts of positive region optimal scale combinations (POSCs) and conditional entropy optimal scale combinations (CEOSCs) in GMHDSs are defined, respectively. Furthermore, algorithms for calculating a POSC and a CEOSC in a GMHDS are also formulated. Finally, a comparative study of our methods with these based on original data under KNN and NB classifiers is conducted on twelve UCI datasets for the assessment of their classification performance. The experimental results indicate that the proposed methods are of better classification performance than the original data based methods in most cases.

Improving inter-session performance via relevant session-transfer for multi-session motor imagery classification

Motor imagery (MI)-based brain-computer interfaces (BCIs) using electroencephalography (EEG) have found practical applications in external device control. However, the non-stationary nature of EEG signals remains to obstruct BCI performance across multiple sessions, even for the same user. In this study, we aim to address the impact of non-stationarity, also known as inter-session variability, on multi-session MI classification performance by introducing a novel approach, the relevant session-transfer (RST) method. Leveraging the cosine similarity as a benchmark, the RST method transfers relevant EEG data from the previous session to the current one. The effectiveness of the proposed RST method was investigated through performance comparisons with the self-calibrating method, which uses only the data from the current session, and the whole-session transfer method, which utilizes data from all prior sessions. We validated the effectiveness of these methods using two datasets: a large MI public dataset (Shu Dataset) and our own dataset of gait-related MI, which includes both healthy participants and individuals with spinal cord injuries. Our experimental results revealed that the proposed RST method leads to a 2.29 % improvement (p < 0.001) in the Shu Dataset and up to a 6.37 % improvement in our dataset when compared to the self-calibrating method. Moreover, our method surpassed the performance of the recent highest-performing method that utilized the Shu Dataset, providing further support for the efficacy of the RST method in improving multi-session MI classification performance. Consequently, our findings confirm that the proposed RST method can improve classification performance across multiple sessions in practical MI-BCIs.

Identification of millet origin using terahertz spectroscopy combined with ensemble learning

It’s crucial for both producers and consumers to accurately trace the origin of millet, given the significant differences in price and taste that exist between millets from various origins. The traditional method of identifying the origin of millet is time-consuming, laborious, complex, and destructive. In this study, a new method for fast and non-destructive differentiation of millet origins is developed by combining terahertz time domain spectroscopy with ensemble learning. Firstly, three machine learning algorithms, namely support vector machine (SVM), random forest (RF), and kernel extreme learning machine (KELM), were used to build different discriminative models, and then the impact of six different preprocessing methods on the models’ classification performance was compared. It was observed that models employing Savitzky-Golay preprocessing exhibited pronounced superiority in accurately determining the millet’s geographical origins. Building upon these findings, the research introduces an innovative ensemble learning strategy, leveraging both topsis and stacking techniques, to harness the collective strengths of the three algorithms. The outcomes of this approach reveal its remarkable capacity to distinguish millets originating from five distinct locations without the necessity for any parameter fine-tuning. The accuracy, F1 score, and Kappa on the prediction set are all 100 %, which significantly outperforms the single model, traditional voting method, and stacking method. The culmination of this study suggests that the integration of terahertz time-domain spectroscopy and TOPSIS-Stacking ensemble learning emerges as a promising method for the swift and non-intrusive discrimination of millet geographical origins with remarkable precision.

Multi-Domain Based Dynamic Graph Representation Learning for EEG Emotion Recognition.

Graph neural networks (GNNs) have demonstrated efficient processing of graph-structured data, making them a promising method for electroencephalogram (EEG) emotion recognition. However, due to dynamic functional connectivity and nonlinear relationships between brain regions, representing EEG as graph data remains a great challenge. To solve this problem, we proposed a multi-domain based graph representation learning (MD 2GRL) framework to model EEG signals as graph data. Specifically, MD 2GRL leverages gated recurrent units (GRU) and power spectral density (PSD) to construct node features of two subgraphs. Subsequently, the self-attention mechanism is adopted to learn the similarity matrix between nodes and fuse it with the intrinsic spatial matrix of EEG to compute the corresponding adjacency matrix. In addition, we introduced a learnable soft thresholding operator to sparsify the adjacency matrix to reduce noise in the graph structure. In the downstream task, we designed a dual-branch GNN and incorporated spatial asymmetry for graph coarsening. We conducted experiments using the publicly available datasets SEED and DEAP, separately for subject-dependent and subject-independent, to evaluate the performance of our model in emotion classification. Experimental results demonstrated that our method achieved state-of-the-art (SOTA) classification performance in both subject-dependent and subject-independent experiments. Furthermore, the visualization analysis of the learned graph structure reveals EEG channel connections that are significantly related to emotion and suppress irrelevant noise. These findings are consistent with established neuroscience research and demonstrate the potential of our approach in comprehending the neural underpinnings of emotion.

Sensor Fault Detection and Classification Using Multi-Step-Ahead Prediction with an Long Short-Term Memoery (LSTM) Autoencoder

The Internet of Things (IoT) is witnessing a surge in sensor-equipped devices. The data generated by these IoT devices serve as a critical foundation for informed decision-making, real-time insights, and innovative solutions across various applications in everyday life. However, data reliability is often compromised due to the vulnerability of sensors to faults arising from harsh operational conditions that can adversely affect the subsequent operations that depend on the collected data. Hence, the identification of anomalies within sensor-derived data holds significant importance in the IoT context. This article proposes a sensor fault detection method using a Long Short-Term Memory autoencoder (LSTM-AE). The AE, trained on normal sensor data, predicts a 20-step window, generating three statistical features via SHapley Additive exPlanations from the estimated steps. These features aid in determining potential faults in the predicted steps using a machine learning classifier. A secondary classifier identifies the type of fault in the sensor signal. Experimentation on two sensor datasets showcases the method’s functionality, achieving fault detection accuracies of approximately 93% and 97%. It is possible to attain a perfect fault classification performance by slightly modifying the feature calculation approach. In a univariate prediction scenario, our proposed approach demonstrates good fault detection and classification performance.

Classification performance analysis of a louver classifier with a tilted sidewall and its application to the separation of unburned carbon from biomass combustion ash

Separating unburned carbon (UC) from biomass combustion ash is necessary to achieve high-quality ash as recyclable raw material and improve energy efficiency by reburning the UC in the furnace. In this study, a novel louver classifier geometry with tilted sidewall was proposed to separate the UC from biomass ash and its classification performance was analyzed via experiments and simulations. The louver classifier with a tilted sidewall provided a smaller 50% cut size and classification accuracy index indicating that the separation performance is improved. The CFD simulations revealed that tilted sidewall in the louver increased gas flow downward velocity in the classification region, increasing the inertia of the particles, and causing more particles to leave gas flow and deposit into the collection box as coarse ash. The UC of classified combustion ash increased with a decrease in the median mass diameter, indicating that the classification process for combustion ash produced coarse ash with a lower UC. The tilted sidewall in the modified louver classifier increased the reduction of unburned carbon (RUC) in ash up to 83.8%. The RUC increased with a cut size of the classified ash which was achieved when a high inlet flow velocity and high blow-up ratio are applied.

Development of a machine learning algorithm for fake news detection

With the extensive technological advancements and expansion, the persistent issues regarding the creation and rapid dissemination of fake news have become a prevalent and recurrent concern. The manipulation of news content has critical repercussions, such as causing public mistrust, fear, harm, and misinformation. Addressing that, this study developed a supervised machine learning algorithm that can accurately classify social media data as fake news. The methodology of the proposed fake news detection model involved five main components: data acquisition from Twitter, data preprocessing, data transformation, model development using Naïve Bayes, decision tree, and support vector machine (SVM) and model evaluation using accuracy, precision, recall and F1-score. The results revealed that decision tree recorded the highest accuracy for both textual data (100%) and metadata (94.54%) and consistently outperformed both Naïve Bayes and SVM in terms of precision, recall, and F1-score metrics, with a score of 100% for the classification of textual data-based datasets. Regarding the metadata-based classification, decision tree also demonstrated excellent performance, with the highest F1-score of 94% for fake news data. Meanwhile, SVM exhibited the highest precision and recall performance for the metadata-based classification. Overall, the application of the decision tree classifier was deemed the most effective in Twitter fake news detection.

Extracting lung cancer staging descriptors from pathology reports: A generative language model approach

BackgroundIn oncology, electronic health records contain textual key information for the diagnosis, staging, and treatment planning of patients with cancer. However, text data processing requires a lot of time and effort, which limits the utilization of these data. Recent advances in natural language processing (NLP) technology, including large language models, can be applied to cancer research. Particularly, extracting the information required for the pathological stage from surgical pathology reports can be utilized to update cancer staging according to the latest cancer staging guidelines. ObjectivesThis study has two main objectives. The first objective is to evaluate the performance of extracting information from text-based surgical pathology reports and determining pathological stages based on the extracted information using fine-tuned generative language models (GLMs) for patients with lung cancer. The second objective is to determine the feasibility of utilizing relatively small GLMs for information extraction in a resource-constrained computing environment. MethodsLung cancer surgical pathology reports were collected from the Common Data Model database of Seoul National University Bundang Hospital (SNUBH), a tertiary hospital in Korea. We selected 42 descriptors necessary for tumor-node (TN) classification based on these reports and created a gold standard with validation by two clinical experts. The pathology reports and gold standard were used to generate prompt-response pairs for training and evaluating GLMs which then were used to extract information required for staging from pathology reports. ResultsWe evaluated the information extraction performance of six trained models as well as their performance in TN classification using the extracted information. The Deductive Mistral-7B model, which was pre-trained with the deductive dataset, showed the best performance overall, with an exact match ratio of 92.24% in the information extraction problem and an accuracy of 0.9876 (predicting T and N classification concurrently) in classification. ConclusionThis study demonstrated that training GLMs with deductive datasets can improve information extraction performance, and GLMs with a relatively small number of parameters at approximately seven billion can achieve high performance in this problem. The proposed GLM-based information extraction method is expected to be useful in clinical decision-making support, lung cancer staging and research.

Deep Learning Features Can Improve Radiomics-Based Prostate Cancer Aggressiveness Prediction

PURPOSE Emerging evidence suggests that the use of artificial intelligence can assist in the timely detection and optimization of therapeutic approach in patients with prostate cancer. The conventional perspective on radiomics encompassing segmentation and the extraction of radiomic features considers it as an independent and sequential process. However, it is not necessary to adhere to this viewpoint. In this study, we show that besides generating masks from which radiomic features can be extracted, prostate segmentation and reconstruction models provide valuable information in their feature space, which can improve the quality of radiomic signatures models for disease aggressiveness classification. MATERIALS AND METHODS We perform 2,244 experiments with deep learning features extracted from 13 different models trained using different anatomic zones and characterize how modeling decisions, such as deep feature aggregation and dimensionality reduction, affect performance. RESULTS While models using deep features from full gland and radiomic features consistently lead to improved disease aggressiveness prediction performance, others are detrimental. Our results suggest that the use of deep features can be beneficial, but an appropriate and comprehensive assessment is necessary to ensure that their inclusion does not harm predictive performance. CONCLUSION The study findings reveal that incorporating deep features derived from autoencoder models trained to reconstruct the full prostate gland (both zonal models show worse performance than radiomics only models), combined with radiomic features, often lead to a statistically significant increase in model performance for disease aggressiveness classification. Additionally, the results also demonstrate that the choice of feature selection is key to achieving good performance, with principal component analysis (PCA) and PCA + relief being the best approaches and that there is no clear difference between the three proposed latent representation extraction techniques.

A Dynamic Multi‐Output Convolutional Neural Network for Skin Lesion Classification

ABSTRACTSkin cancer is a pressing global health issue, with high incidence and mortality rates. Convolutional neural network (CNN) models have been proven to be effective in improving performance in skin lesion image classification and reducing the medical burden. However, the inherent class imbalance in training data caused by the difficulty of collecting dermoscopy images leads to categorical overfitting, which still limits the effectiveness of these data‐driven models in recognizing few‐shot categories. To address this challenge, we propose a dynamic multi‐output convolutional neural network (DMO‐CNN) model that incorporates exit nodes into the standard CNN structure and includes feature refinement layers (FRLs) and an adaptive output scheduling (AOS) module. This model improves feature representation ability through multi‐scale sub‐feature maps and reduces the inter‐layer dependencies during backpropagation. The FRLs ensure efficient and low‐loss down‐sampling, while the AOS module utilizes a trainable layer selection mechanism to refocus the model's attention on few‐shot lesion categories. Additionally, a novel correction factor loss is introduced to supervise and promote AOS convergence. Our evaluation of the DMO‐CNN model on the HAM10000 dataset demonstrates its effectiveness in multi‐class skin lesion classification and its superior performance in recognizing few‐shot categories. Despite utilizing a very simple VGG structure as the sole backbone structure, DMO‐CNN achieved impressive performance of 0.885 in BACC and 0.983 in weighted AUC. These results are comparable to those of the ensemble model that won the ISIC 2018 challenge, highlighting the strong potential of DMO‐CNN in dealing with few‐shot skin lesion data.

FBENet: Feature-Level Boosting Ensemble Network for Hashimoto's Thyroiditis Ultrasound Image Classification.

Distinguishing Hashimoto's thyroiditis (HT) lesions from ordinary thyroid tissues is difficult with ultrasound images. Challenges in achieving high performance of HT ultrasound image classification include the low resolution, blurred features and large area of irrelevant noise. To address these problems, we propose a Feature-level Boosting Ensemble Network (FBENet) for HT ultrasound image classification. Specifically, to capture the features of suspicious HT lesions efficiently, an Ensemble Feature Boosting Module (EFBM) is introduced into the feature-level ensemble to boost the blurred features. Then, the spatial attention mechanism is adopted in backbone models to improve the feature focusing performance and representation ability. Furthermore, feature-level ensemble technique is employed in the training process to achieve more comprehensive feature representation ability. Experimentally, FBENet was trained on 6,503 HT ultrasound images, and tested on 1,626 HT ultrasound images with 82.92% accuracy and 89.24% AUC on average.