Fitri Arnia

Cross-spectral cross-distance face recognition (CSCD-FR) is a challenging biometric task involving visible (VIS) and near-infrared (NIR) facial images acquired at varying stand-off distances. Beyond the spectral discrepancy between VIS and NIR images, increasing acquisition distance degrades resolution, blurs images, and reduces facial detail, making recognition considerably more difficult. Although deep convolutional neural networks (CNNs) have demonstrated strong capabilities for learning discriminative facial embeddings, existing studies have predominantly focused on network architectures, feature representations, and learning strategies, while the influence of distance metric selection remains relatively underexplored. Since recognition decisions are ultimately based on distances between embedding vectors, the choice of similarity metric may affect recognition performance. This study therefore investigates the influence of generalized Minkowski distance order on CSCD-FR using deep embeddings extracted from MobileNetV3-Small, EfficientNet-B2, and DenseNet121. Using the Long-Distance Heterogeneous Face Database (LDHF-DB) at stand-off distances of 1, 60, and 100 m, both ImageNet-pretrained and domain-adapted models were evaluated using -normalized embeddings. Domain adaptation optimized the embedding layer using triplet margin loss while the CNN backbones remained frozen. Minkowski orders from to were evaluated using closed-set identification and verification protocols. ImageNet-pretrained models yielded Rank-1 accuracies below 11%, whereas domain adaptation substantially improved recognition performance. DenseNet121 achieved the highest observed performance, with 93.19% Rank-1 accuracy and 2.16% EER at . The Rank-1-optimal Minkowski distance order varied across architectures, with EfficientNet-B2 and MobileNetV3-Small achieving their highest observed Rank-1 accuracies at and , respectively. These findings show that selecting a similarity metric can provide an additional matching-stage optimization without modifying the CNN backbone. Because the same retained identities were used for domain adaptation and evaluation, the results represent a dataset-specific assessment rather than recognition performance on previously unseen subjects.

Objective pain assessment in non-verbal patients remains a significant clinical challenge. While automated facial expression analysis offers a promising solution, existing methods often rely on simple cross-modality concatenation or shallow attention mechanisms, which typically fail to fully capture complex, interdependent cross-modal dynamics. To address this limitation, this study proposes a deep learning-based multimodal feature fusion framework that combines visual and thermal images to significantly enhance pain detection accuracy. The proposed framework integrates pretrained convolutional neural network backbones, specifically VGGFace, ResNet50, and DEYOLO, with two novel attention modules: Dual Semantic Enhancing Channel Weight Assignment (DECA), and Dual Spatial Enhancing Pixel Weight Assignment (DEPA) to adaptively optimize joint feature representations. For comparison, a hybrid baseline model that processes visual and thermal images separately was also developed, allowing a comparative analysis between fusion-based and non-fusion-based approaches. The model performance was systematically evaluated on the MIntPain dataset, which comprises 20 healthy subjects experiencing five distinct levels of pain intensity. To ensure data independence and prevent identity leakage across the training and testing phases, a strict subject-wise data split protocol was implemented with 15 subjects allocated for training and 5 subjects for testing. Experimental results demonstrate that the proposed multimodal fusion framework achieves superior performance, attaining a peak F1 score of 0.938 using the VGGFace backbone. Furthermore, external validation on the UNBC McMaster Shoulder Pain dataset yields a classification accuracy of 0.872, confirming the strong generalization capability and stability of the framework across unseen subjects. These findings highlight the effectiveness of visual-thermal synergy and the efficacy of the proposed DECA and DEPA modules, showcasing high potential for robust, non-invasive clinical pain-monitoring and assessment applications

Individual classification models for caries detection still face significant challenges, including limited accuracy and unstable predictions, which can hinder diagnosis, delay clinical decisions, and increase the risks associated with patient care. To overcome these limitations, this study proposes an ensemble voting method that combines five deep learning models, such as ResNet-152, MobileNetV2, InceptionV3, NASNetMobile, and EfficientNet-B5. This approach aims to enhance the accuracy and stability of caries detection by leveraging the complementary strengths of the individual models while mitigating their weaknesses. Each model was trained and tested on the same dataset of dental images, categorized into caries and regular classes. Their predictions were aggregated using hard and soft voting techniques. The ensemble's performance was evaluated using accuracy, precision, recall, and F1-score. The ensemble voting demonstrates a notable improvement in classification performance over individual models. Hard and soft voting have excellent classification performance and consistently outperform the best individual models. The accuracy increased from EfficientNetB5 0.8485 to 0.8864 and 0.8712, representing increases of 4.46% and 2.68%, respectively. The precision increased from MobileNetV2 0.8182 to 0.8493 and 0.8551, representing increases of 3.81% and 4.52%. For recall, EfficientNetB5 ranked highest among individual models with a score of 0.9242. Hard voting increased 1.64% to 0.9394, and soft voting decreased slightly by 3.28% to 0.8939. The F1 score of EfficientNetB5 is 0.8592. Hard and soft voting increased 3.83% and 1.73% to 0.8921 and 0.8741. The proposed ensemble improves the F1-score by 3.83 percentage points compared to the best individual model. The ensemble voting method effectively leverages the complementary strengths of each deep learning model to improve the stability and accuracy of fast, reliable dental caries early detection prediction.

Tuberculosis (TB) remains a major global health challenge that requires diagnostic solutions capable of integrating structural and functional information from the respiratory system. This study proposes a novel lesion‐guided multimodal framework that fuses chest X‐ray (CXR) and lung sound data for TB detection, addressing the lack of synchronized datasets by leveraging a data‐driven pairing mechanism. A YOLOv8‐seg model was employed to localize pulmonary lesions, which are then mapped to anatomically corresponding auscultation sites using a rib‐guided quadrant extraction algorithm. This process generates a physiologically coherent paired dataset containing 450 CXR images and 2700 lung sound segments. A dual‐branch architecture is developed, consisting of a ResNet50 encoder for visual features and a convolutional recurrent neural network (CRNN) encoder for mel‐frequency cepstral coefficient (MFCC)–based acoustic representations. Their probabilistic outputs are integrated using an optimized weighted decision‐level fusion strategy, with weights derived from unimodal validation performance. Simulation results show that the proposed multimodal framework outperforms both unimodal baselines, achieving 92% accuracy and an AUC of 0.98, compared with 0.95 (CXR‐only) and 0.81 (audio‐only). Qualitative Grad‐CAM and saliency analyses further confirm that the model focuses on interpretable radiographic lesions and crackle‐like audio patterns. These findings demonstrate that the proposed lesion‐guided multimodal framework offers a physiologically consistent and data‐efficient approach for TB screening, highlighting its potential for future clinically deployable diagnostic systems.

Several studies have been conducted to develop tuberculosis detection using artificial intelligence-based chest radiographs (CXR). However, CXR images also contain atypical information on pulmonary tuberculosis that could affect the accuracy of the learning. This work aims to identify the regions of interest (RoI) in CXR images that support information on secondary pulmonary tuberculosis in the upper lung area and train a vision transformer network to build a robust CXR classifier. The dataset used in this study consists of 1,304 annotated CXR images. The ground-truth masks of the upper one-third lung region were manually created using ImageJ. The methods include enhancing the CXR image using Contrast-limited Adaptive Histogram Equalization before extracting RoI using the Attention U-Net segmentation algorithm. The informative CXR images were then fed into the self-attention network of the vision transformer (ViT-L/16). The results show that the segmentation model achieves 100% accuracy in identifying important areas on images that have undergone image contrast pre-processing. The classifier model exhibits stable performance during training, achieving an accuracy of 97.75%. The test results on 260 images, 130 normal and 130 TB, showthat TBNet achieves 97.04% accuracy, 97.96% sensitivity, 96.92% specificity, and 96.94% precision. TBNet achieves a high accuracy rate and has a size of 127 MB for segmentation and 1.12 GB for classification; thus, practically, the proposed TBNet can be deployed to support tuberculosis detection system. These findings contribute to research on medical image analysis by highlighting the importance of region-specific annotation and transformer-based architectures for secondary tuberculosis detection.

The subject matter is that the characteristics of tuberculosis are difficult to study visually. Therefore, a computer-aided system based on deep learning can be applied to X-ray image recognition. Many studies have been conducted in this area but have yet to achieve a high accuracy rate. The goal of this study is to determine the effect of using various datasets in developing deep learning models. The tasks to be solved include exploring various deep learning architectures and deep fine-tuning hyperparameters, as well as using various dataset sources. The method used is the development of a deep learning model of convolutional neural network (CNN) using transfer learning to classify X-ray images into binary classes of normal and tuberculosis (TB). The CNN architectures used are the pretrained networks of ResNet and EfficientNet, along with their variants. The pre-trained network was trained on a dataset obtained from four sources: Shenzhen, Montgomery, RSNA CXR, and Belarus. The dataset is divided into three schemes: Scheme one consists of the Shenzhen dataset with low-quality X-ray images; Scheme two is the Montgomery, RSNA, and Belarus datasets that show good contrast in the indicated TB area; and Scheme three contains datasets from all sources to allow for more datasets to be learned. The augmentation, dropout, and L2 regularization methods were also applied to enhance learning performance. The following results were obtained: the models performed better with the high-quality X-ray images in Scheme Two but not with the large dataset in Scheme Three. Regarding network performance, the models resulting from ResNet-101 and EfficientNetB0 outperformed the others with good fit learning and capability in recognizing X-ray images with an accuracy rate of 99.2%. In conclusion, the best approach to enhance learning performance is to use high-quality input and apply regularizations.

It would be more beneficial and practical to detect obesity using typical regions of body thermograms instead of all regions of thermograms. In this study, we aimed to (1) determine the typical region of body thermograms and (2) propose a soft‐vote fusion algorithm to develop a more accurate result in obesity detection. The typical regions of thermograms were selected by training and testing the pretrained CNN models with the thermogram dataset of all‐region and single‐region. All‐region dataset includes thermograms from five regions, namely, supraclavicular (SCV), abdomen, forearm, palm, and shank, while each of single‐region dataset consists of thermograms from an individual region. When tested on all‐region dataset, the CNN model achieved the highest accuracy of 90.00%. When tested with a single‐region dataset, three regions, namely, SCV, abdomen, and forearm, achieved the highest accuracy of 94.23% (SCV) and 96.15% (abdomen and forearm), which indicated that these regions were more representative than palm and shank. Then, we proposed a soft‐vote fusion algorithm in which different trained CNN models from the previous experiments were combined to improve the detection performance. We found that the combination of those three regions can achieve the highest accuracy with the smallest model size. One of those proposed combinations achieved an accuracy of 98.08% with a framework size of 19.4 MB by only using SCV and abdomen thermogram as the inputs.

The image quality supports a high accuracy rate of medical image diagnosis using computer vision. Digital thermal images resulting from the thermal device usually suffer from many watermarks that may lower the neural network learning performance. Thus, providing only the region of interest (RoI) of the breast area from the breast thermal images for early breast cancer detection is an important task. The goal of our work area to develop a deep learning (DL) model for taking the RoI of the breast thermal images, built a self-supervised DL model to classify the breast thermal images into healthy and cancer categories, and integrated these two models as end-to-end bi-pipeline model for breast thermal image recognition. The segmentation model was built using attention U-Net with residual recurrent network called R2AU-Net, and the classification model was built using self-supervised learning consisting of the Simple Framework for Contrastive Learning of Visual Representations (SimCLR) and ResNet50. These networks were trained using unlabelled limited breast thermal datasets to allow more comprehensive learning. The result shows that proposed self-supervised bi-pipeline model can take the RoI with an accuracy of 98.63% and classify the breast thermal images with a top-1 accuracy rate of 84.37% and top-5 accuracy of 96.87%. In addition, the bi-pipeline model implementation using a central processing unit shows that the model required only about 4 seconds for segmentation and classification tasks. These findings indicate that the bi-pipeline model can effectively aid the interpretation of unlabeled breast thermal images.

Classification of stunted and normal children using novel facial image database and convolutional neural network