Kusworo Adi
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Kusworo Adi has a total of 97 co-authors. Here are the top co-authors by number of shared publications:
- Catur Edi Widodo
Diponegoro University
13 shared publications - Rahmat Gernowo
Universitas Diponegoro
7 shared publications Adi Pamungkas
Department of Physics, Diponegoro University, Semarang, Indonesia
5 shared publicationsCatur Edi Widodo
5 shared publicationsOky Dwi Nurhayati
Department of Computer System, Diponegoro University, Semarang, Indonesia
4 shared publicationsSri Pujiyanto
Department of Biology, Diponegoro University, Semarang, Indonesia
4 shared publications
Latest Publications
- Open Access
- 1 October 2026
- Published by Institute of Advanced Engineering and Science in IAES International Journal of Artificial Intelligence (IJ-AI)
- Vol. 15 (5) , 4121-4132
Automated convolutional neural network-based reject–repeat decision system for chest radiographs
Image quality in radiography is essential for diagnostic accuracy, yet reject–repeat analysis (RRA) is still mostly performed manually and depends on subjective radiographer judgment. Inter-operator variability leads to inconsistent quality data and limits workflow optimization. While many studies evaluate deep learning models for disease classification, research using convolutional neural networks (CNNs) to assess image quality and automate RRA remains limited. This study introduces a novel approach by applying CNNs to classify the causes of image rejection in chest radiography: positioning, exposure, artefacts, and accepted images, integrating these outcomes into a quality-management framework. The aim is to develop and evaluate CNN models (visual geometry group (VGG)-16, residual neural network (ResNet)-101, and dense convolutional network (DenseNet)-201) to support automated RRA using digital imaging and communications in medicine (DICOM)-based radiographs. The dataset comprised DICOM chest images annotated by a panel of radiologists, radiographers, and medical physicists into four technical classes. The CNNs were trained using supervised learning with structured training, validation, and testing splits, and evaluated using accuracy, precision, recall, and F1-score. Results show VGG-16 delivers the most stable and accurate performance across categories. Overall, CNNs effectively automate RRA with greater objectivity and consistency, reducing repeat rates, improving workflow efficiency, and strengthening radiology quality-management systems to enhance patient safety.
- 1 October 2026
- Published by Institute of Advanced Engineering and Science in IAES International Journal of Artificial Intelligence (IJ-AI)
- Vol. 15 (5) , 4121-4132
Open AccessAutomated convolutional neural network-based reject–repeat decision system for chest radiographs
Image quality in radiography is essential for diagnostic accuracy, yet reject–repeat analysis (RRA) is still mostly performed manually and depends on subjective radiographer judgment. Inter-operator variability leads to inconsistent quality data and limits workflow optimization. While many studies evaluate deep learning models for disease classification, research using convolutional neural networks (CNNs) to assess image quality and automate RRA remains limited. This study introduces a novel approach by applying CNNs to classify the causes of image rejection in chest radiography: positioning, exposure, artefacts, and accepted images, integrating these outcomes into a quality-management framework. The aim is to develop and evaluate CNN models (visual geometry group (VGG)-16, residual neural network (ResNet)-101, and dense convolutional network (DenseNet)-201) to support automated RRA using digital imaging and communications in medicine (DICOM)-based radiographs. The dataset comprised DICOM chest images annotated by a panel of radiologists, radiographers, and medical physicists into four technical classes. The CNNs were trained using supervised learning with structured training, validation, and testing splits, and evaluated using accuracy, precision, recall, and F1-score. Results show VGG-16 delivers the most stable and accurate performance across categories. Overall, CNNs effectively automate RRA with greater objectivity and consistency, reducing repeat rates, improving workflow efficiency, and strengthening radiology quality-management systems to enhance patient safety.
- Open Access
- 24 July 2026
- Published by International Academic Press in Statistics, Optimization & Information Computing
- Vol. 16 (4) , 3252-3271
Benchmarking Attention Mechanisms in ConvNeXt-Tiny for Dermoscopic Skin Lesion Classification Under Stratified and Lesion-Level Leakage-Safe Cross-Validation
Attention mechanisms are widely added to convolutional backbones to improve skin lesion classification, yet their comparative benefit is rarely assessed under evaluation protocols that prevent lesion-level data leakage. We present a controlled benchmark of four canonical attention modules—Channel Attention, Spatial Attention, the Convolutional Block Attention Module (CBAM), and Coordinate Attention—each inserted as a post-backbone plug-in on a shared ConvNeXt-Tiny feature extractor for seven-class dermoscopic classification on the HAM10000 dataset. Every configuration is evaluated twice: under conventional image-level stratified five-fold cross-validation, and under lesion-aware group five-fold cross-validation that ensures all images sharing the same lesion id remain within a single fold. Across all configurations, replacing stratified with group validation reduces the macro-averaged F1-score by an average of 10.82 percentage points, an inflation concentrated mainly in the minority classes—the rarest class loses more than 22 points of F1—and shows an exploratory negative association with class rarity (r = −0.66). The ranking of attention mechanisms is not preserved across protocols: Channel Attention attains the highest observed macro-F1 under image-level stratified validation, whereas Spatial Attention attains the highest observed macro-F1 under lesion-level group validation and is the only attention mechanism with a numerically higher mean than the corresponding attention-free baseline. However, none of the eight fold-level comparisons against the protocol-matched baselines remains statistically significant after Holm correction. Peak GPU-memory requirements are broadly comparable across configurations, although Coordinate Attention requires moderately longer training under lesion-level group validation. These results show that the cross-validation protocol influences reported performance more strongly than the choice of attention mechanism, and that image-level evaluation on HAM10000 can provide optimistic estimates of generalisation to unseen lesions. We recommend lesion-level group partitioning as the preferred protocol for trustworthy evaluation of skin lesion classifiers.
- 24 July 2026
- Published by International Academic Press in Statistics, Optimization & Information Computing
- Vol. 16 (4) , 3252-3271
Open AccessBenchmarking Attention Mechanisms in ConvNeXt-Tiny for Dermoscopic Skin Lesion Classification Under Stratified and Lesion-Level Leakage-Safe Cross-Validation
Attention mechanisms are widely added to convolutional backbones to improve skin lesion classification, yet their comparative benefit is rarely assessed under evaluation protocols that prevent lesion-level data leakage. We present a controlled benchmark of four canonical attention modules—Channel Attention, Spatial Attention, the Convolutional Block Attention Module (CBAM), and Coordinate Attention—each inserted as a post-backbone plug-in on a shared ConvNeXt-Tiny feature extractor for seven-class dermoscopic classification on the HAM10000 dataset. Every configuration is evaluated twice: under conventional image-level stratified five-fold cross-validation, and under lesion-aware group five-fold cross-validation that ensures all images sharing the same lesion id remain within a single fold. Across all configurations, replacing stratified with group validation reduces the macro-averaged F1-score by an average of 10.82 percentage points, an inflation concentrated mainly in the minority classes—the rarest class loses more than 22 points of F1—and shows an exploratory negative association with class rarity (r = −0.66). The ranking of attention mechanisms is not preserved across protocols: Channel Attention attains the highest observed macro-F1 under image-level stratified validation, whereas Spatial Attention attains the highest observed macro-F1 under lesion-level group validation and is the only attention mechanism with a numerically higher mean than the corresponding attention-free baseline. However, none of the eight fold-level comparisons against the protocol-matched baselines remains statistically significant after Holm correction. Peak GPU-memory requirements are broadly comparable across configurations, although Coordinate Attention requires moderately longer training under lesion-level group validation. These results show that the cross-validation protocol influences reported performance more strongly than the choice of attention mechanism, and that image-level evaluation on HAM10000 can provide optimistic estimates of generalisation to unseen lesions. We recommend lesion-level group partitioning as the preferred protocol for trustworthy evaluation of skin lesion classifiers.
- Open Access
- 1 July 2026
- Published by Society of Visual Informatics in JOIV : International Journal on Informatics Visualization
- Vol. 10 (4) , 1548-1554
A Deep Learning Methodology for the Early Identification of Chili Plant Diseases Utilizing SSD-ResNet-50 Architecture
The prompt identification and precise categorization of chili plant diseases are crucial for promoting sustainable agriculture and reducing crop losses due to pests and pathogens. This research presents an automated detection method using the Single Shot Detector (SSD) framework, with ResNet-50 as the backbone and a Feature Pyramid Network (FPN) to manage multi-scale feature representations. Dataset including 14,248 images of chili plants, categorized into six classes like Leaf Spot, Rotten Fruit, Healthy Fruit, Healthy Leaf, Mosaic Curl, and Yellowing, underwent preprocessing involving segmentation, resizing, and augmentation, followed by a division into 90% training data and 10% testing data. Transfer learning was implemented using a COCO-pretrained SSD ResNet-50 FPN model, enhanced with cosine-decay learning-rate scheduling and momentum optimization. The assessment results indicated an overall accuracy of 92.7%, with the highest F1-scores achieved for Healthy Fruit (0.965) and Rotten Fruit (0.967). Under the COCO evaluation protocol, the model achieved an mAP@0.5 of 91.5% and mAP@[0.5:0.95] of 65.2%. Model ran at approximately 30 FPS on an NVIDIA T4 GPU. Reduced precision values were noted for Leaf Spot (0.866) and Mosaic Curl (0.850), suggesting a propensity for misclassification due to visual similarities among disease symptoms. Nonetheless, all classes attained F1-scores exceeding 0.86, illustrating the robustness of the proposed model. Importantly, the SSD-ResNet-50 approach offers both efficiency and accuracy within a single pipeline, enabling rapid inference practical for real-world applications. These findings emphasize the potential of deep learning-based solutions to strengthen plant disease monitoring systems. In conclusion, SSD with ResNet-50 FPN provides an effective and scalable methodology for the automated identification of chili plant diseases, contributing directly to sustainable agriculture and improved crop management practices.
- 1 July 2026
- Published by Society of Visual Informatics in JOIV : International Journal on Informatics Visualization
- Vol. 10 (4) , 1548-1554
Open AccessA Deep Learning Methodology for the Early Identification of Chili Plant Diseases Utilizing SSD-ResNet-50 Architecture
The prompt identification and precise categorization of chili plant diseases are crucial for promoting sustainable agriculture and reducing crop losses due to pests and pathogens. This research presents an automated detection method using the Single Shot Detector (SSD) framework, with ResNet-50 as the backbone and a Feature Pyramid Network (FPN) to manage multi-scale feature representations. Dataset including 14,248 images of chili plants, categorized into six classes like Leaf Spot, Rotten Fruit, Healthy Fruit, Healthy Leaf, Mosaic Curl, and Yellowing, underwent preprocessing involving segmentation, resizing, and augmentation, followed by a division into 90% training data and 10% testing data. Transfer learning was implemented using a COCO-pretrained SSD ResNet-50 FPN model, enhanced with cosine-decay learning-rate scheduling and momentum optimization. The assessment results indicated an overall accuracy of 92.7%, with the highest F1-scores achieved for Healthy Fruit (0.965) and Rotten Fruit (0.967). Under the COCO evaluation protocol, the model achieved an mAP@0.5 of 91.5% and mAP@[0.5:0.95] of 65.2%. Model ran at approximately 30 FPS on an NVIDIA T4 GPU. Reduced precision values were noted for Leaf Spot (0.866) and Mosaic Curl (0.850), suggesting a propensity for misclassification due to visual similarities among disease symptoms. Nonetheless, all classes attained F1-scores exceeding 0.86, illustrating the robustness of the proposed model. Importantly, the SSD-ResNet-50 approach offers both efficiency and accuracy within a single pipeline, enabling rapid inference practical for real-world applications. These findings emphasize the potential of deep learning-based solutions to strengthen plant disease monitoring systems. In conclusion, SSD with ResNet-50 FPN provides an effective and scalable methodology for the automated identification of chili plant diseases, contributing directly to sustainable agriculture and improved crop management practices.
- Open Access
- 18 April 2026
- Published by Elsevier in SSRN Electronic Journal
Cooperative encirclement strategy of swarm unmanned surface vehicles via Tesla-valve optimization-based multi-agent reinforcement learning (TVO-MARL)
Swarm unmanned surface vehicles (USVs) have become significantly prominent for maritime defense and security operations. Cooperative encirclement task, as part of maritime operations, is a critical capability of autonomous swarm USVs that requires motion stability, robust coordination, and energy efficiency under dynamic conditions. The key challenges in encirclement tasks are that most approaches focus solely on translational and positional motion while neglecting the integration of rotational motion around the target. In addition, convergence speed remains a persistent issue in reinforcement learning-based methods. To address these challenges, this work proposes the integration of Tesla-valve optimization into multi-agent reinforcement learning (TVO-MARL) to improve the encirclement task performance. The novel TVO approach is applied into two baselines of MARL including multi-agent proximal policy optimization (MAPPO), and MAPPO-LSTM (long short-term memory). The proposed methods are assessed in the 3D simulation platform using multiple performance metrics, including cumulative reward, angular coverage, rotation number, coverage variance, trajectory smoothness, and energy consumption. Experimental results show that proposed methods outperform the baseline for the convergence speed in training stage. Moreover, the top-performing trained models demonstrate that MAPPO-TVO significantly improves rotational stability, while MAPPO-LSTM-TVO achieves the highest cumulative reward and angular coverage, along with smoother trajectories and substantially reduced energy consumption. Statistical analysis confirms that these improvements are significant compared to baseline MAPPO and MAPPO-LSTM methods. The results reveal clear trade-offs across encirclement objectives and highlight the complementary benefits of nature-inspired optimization via TVO and temporal policy representation, providing a robust solution for swarm USV encirclement task.
- 18 April 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessCooperative encirclement strategy of swarm unmanned surface vehicles via Tesla-valve optimization-based multi-agent reinforcement learning (TVO-MARL)
Swarm unmanned surface vehicles (USVs) have become significantly prominent for maritime defense and security operations. Cooperative encirclement task, as part of maritime operations, is a critical capability of autonomous swarm USVs that requires motion stability, robust coordination, and energy efficiency under dynamic conditions. The key challenges in encirclement tasks are that most approaches focus solely on translational and positional motion while neglecting the integration of rotational motion around the target. In addition, convergence speed remains a persistent issue in reinforcement learning-based methods. To address these challenges, this work proposes the integration of Tesla-valve optimization into multi-agent reinforcement learning (TVO-MARL) to improve the encirclement task performance. The novel TVO approach is applied into two baselines of MARL including multi-agent proximal policy optimization (MAPPO), and MAPPO-LSTM (long short-term memory). The proposed methods are assessed in the 3D simulation platform using multiple performance metrics, including cumulative reward, angular coverage, rotation number, coverage variance, trajectory smoothness, and energy consumption. Experimental results show that proposed methods outperform the baseline for the convergence speed in training stage. Moreover, the top-performing trained models demonstrate that MAPPO-TVO significantly improves rotational stability, while MAPPO-LSTM-TVO achieves the highest cumulative reward and angular coverage, along with smoother trajectories and substantially reduced energy consumption. Statistical analysis confirms that these improvements are significant compared to baseline MAPPO and MAPPO-LSTM methods. The results reveal clear trade-offs across encirclement objectives and highlight the complementary benefits of nature-inspired optimization via TVO and temporal policy representation, providing a robust solution for swarm USV encirclement task.
- Open Access
- 31 December 2025
- Published by Politechnika Lubelska in Applied Computer Science
- Vol. 21 (4) , 1-12
Real-time detection of seat belt usage in overhead traffic surveillance using YOLOv7
Driving safety plays a critical role in minimizing traffic accidents, and seat belt usage is one of the most effective preventive measures. This study aims to implement the YOLOv7 object detection model to automatically detect seat belt usage in four-wheeled vehicles using overhead traffic surveillance images. The proposed method consists of three main stages: dataset preparation, model training, and model evaluation. Dataset preparation includes acquiring video footage from different locations and time conditions, extracting image frames, and annotating four object classes: car, windshield, passenger, and seat belt. The model is trained on a dataset consisting of images taken during both day and night conditions. During training, data augmentation and anchor box optimization are applied to improve model generalization. The trained model is evaluated on an unseen test dataset and achieves a Mean Average Precision at 50% Intersection over Union (mAP50) of 97.46% and an F1 score of 95.37% at the optimal confidence level. These results indicate high detection accuracy for all object classes, especially for the seat belt class with an AP of 93.40%. The proposed system offers a promising solution for real-time traffic enforcement, reducing the reliance on manual observation and potentially improving traffic safety monitoring.
- 31 December 2025
- Published by Politechnika Lubelska in Applied Computer Science
- Vol. 21 (4) , 1-12
Open AccessReal-time detection of seat belt usage in overhead traffic surveillance using YOLOv7
Driving safety plays a critical role in minimizing traffic accidents, and seat belt usage is one of the most effective preventive measures. This study aims to implement the YOLOv7 object detection model to automatically detect seat belt usage in four-wheeled vehicles using overhead traffic surveillance images. The proposed method consists of three main stages: dataset preparation, model training, and model evaluation. Dataset preparation includes acquiring video footage from different locations and time conditions, extracting image frames, and annotating four object classes: car, windshield, passenger, and seat belt. The model is trained on a dataset consisting of images taken during both day and night conditions. During training, data augmentation and anchor box optimization are applied to improve model generalization. The trained model is evaluated on an unseen test dataset and achieves a Mean Average Precision at 50% Intersection over Union (mAP50) of 97.46% and an F1 score of 95.37% at the optimal confidence level. These results indicate high detection accuracy for all object classes, especially for the seat belt class with an AP of 93.40%. The proposed system offers a promising solution for real-time traffic enforcement, reducing the reliance on manual observation and potentially improving traffic safety monitoring.
- Open Access
- 10 November 2025
- Published by Acadlore Publishing Services Limited in International Journal of Transport Development and Integration
- Vol. 9 (4) , 688-699
Real-Time Driver Drowsiness Detection Using ViViT for In-Vehicle Monitoring Systems Under Diverse Driving Conditions
This study proposes a novel approach to driver drowsiness detection using the Video Vision Transformer (ViViT) model, which captures both spatial and temporal dynamics simultaneously to analyze eye conditions and head movements. The National Tsing Hua University Driver Drowsiness Detection (NTHU-DDD) dataset, which consists of 36,000 annotated video clips, was utilized for both training and evaluation. The ViViT model is compared to traditional Convolutional Neural Network (CNN) and Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) models, demonstrating superior performance with 96.2% accuracy and 95.9% F1-Score, while maintaining a 28.9 ms/frame inference time suitable for real-time deployment. The ablation study indicates that integrating spatial and temporal attention yields a notable improvement in model accuracy. Furthermore, positional encoding proves essential in preserving spatial coherence within video-based inputs. The model's resilience was tested across a range of challenging conditions including low-light settings, partial occlusions, and drastic head movements and it consistently maintained reliable performance. With a compact footprint of just 89 MB, the ViViT model has been fine-tuned for deployment on embedded platforms such as the Jetson Nano, making it well-suited for edge AI applications. These findings highlight ViViT's promise as a practical and high-performing solution for real-time driver drowsiness detection in real-world scenarios.
- 10 November 2025
- Published by Acadlore Publishing Services Limited in International Journal of Transport Development and Integration
- Vol. 9 (4) , 688-699
Open AccessReal-Time Driver Drowsiness Detection Using ViViT for In-Vehicle Monitoring Systems Under Diverse Driving Conditions
This study proposes a novel approach to driver drowsiness detection using the Video Vision Transformer (ViViT) model, which captures both spatial and temporal dynamics simultaneously to analyze eye conditions and head movements. The National Tsing Hua University Driver Drowsiness Detection (NTHU-DDD) dataset, which consists of 36,000 annotated video clips, was utilized for both training and evaluation. The ViViT model is compared to traditional Convolutional Neural Network (CNN) and Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) models, demonstrating superior performance with 96.2% accuracy and 95.9% F1-Score, while maintaining a 28.9 ms/frame inference time suitable for real-time deployment. The ablation study indicates that integrating spatial and temporal attention yields a notable improvement in model accuracy. Furthermore, positional encoding proves essential in preserving spatial coherence within video-based inputs. The model's resilience was tested across a range of challenging conditions including low-light settings, partial occlusions, and drastic head movements and it consistently maintained reliable performance. With a compact footprint of just 89 MB, the ViViT model has been fine-tuned for deployment on embedded platforms such as the Jetson Nano, making it well-suited for edge AI applications. These findings highlight ViViT's promise as a practical and high-performing solution for real-time driver drowsiness detection in real-world scenarios.
- Open Access
- 13 October 2025
- Published by LP2M Universitas Islam Negeri (UIN) Syarif Hidayatullah Jakarta in JURNAL TEKNIK INFORMATIKA
- Vol. 18 (2) , 270-280
IMU-Based Early Warning System for Driver Drowsiness Detection via Head Movement Analysis
The high incidence of road accidents caused by human error—accounting for approximately 69.7% of all motor vehicle accidents in Indonesia—demonstrates the urgent need for an effective driver monitoring system. One critical factor contributing to human error is driver drowsiness, which can be observed through behavioral indicators such as abrupt changes in head position. This study aims to develop a real-time early warning system for detecting driver drowsiness based on head movement patterns using a wearable device equipped with the MPU-6050 GY-521 accelerometer sensor. The system monitors acceleration on the X, Y, and Z axes and identifies drowsiness when simultaneous changes exceed predefined thresholds. A drowsiness event is characterized by a rapid head displacement, occurring within approximately 18–20 milliseconds. The thresholds applied for detection are 1.0g for the X axis, 3.5g for the Y axis, and 0.5g for the Z axis. In ten test scenarios simulating drowsy head movements, the system successfully identified seven instances, resulting in a detection accuracy of 70%. The novelty of this approach lies in its lightweight, non-intrusive design and its ability to function independently of lighting conditions, making it a practical solution for real-time driver safety enhancement.
- 13 October 2025
- Published by LP2M Universitas Islam Negeri (UIN) Syarif Hidayatullah Jakarta in JURNAL TEKNIK INFORMATIKA
- Vol. 18 (2) , 270-280
Open AccessIMU-Based Early Warning System for Driver Drowsiness Detection via Head Movement Analysis
The high incidence of road accidents caused by human error—accounting for approximately 69.7% of all motor vehicle accidents in Indonesia—demonstrates the urgent need for an effective driver monitoring system. One critical factor contributing to human error is driver drowsiness, which can be observed through behavioral indicators such as abrupt changes in head position. This study aims to develop a real-time early warning system for detecting driver drowsiness based on head movement patterns using a wearable device equipped with the MPU-6050 GY-521 accelerometer sensor. The system monitors acceleration on the X, Y, and Z axes and identifies drowsiness when simultaneous changes exceed predefined thresholds. A drowsiness event is characterized by a rapid head displacement, occurring within approximately 18–20 milliseconds. The thresholds applied for detection are 1.0g for the X axis, 3.5g for the Y axis, and 0.5g for the Z axis. In ten test scenarios simulating drowsy head movements, the system successfully identified seven instances, resulting in a detection accuracy of 70%. The novelty of this approach lies in its lightweight, non-intrusive design and its ability to function independently of lighting conditions, making it a practical solution for real-time driver safety enhancement.
- Open Access
- 1 February 2025
- Published by Institute of Advanced Engineering and Science in IAES International Journal of Artificial Intelligence (IJ-AI)
- Vol. 14 (1) , 602-613
Utilization of convolutional neural network in image interpretation techniques for detecting kidney disease
This research is conducted with deep learning for kidney stone disease detection including cysts, stones, normal, and tumors using axial computerized tomography (CT) scan images. The author uses augmentation, generative adversarial networks (GANs), original, and synthetic minority over-sampling technique (SMOTE) to classify kidney disease (cyst, stone, normal, and tumor). This study uses the public dataset nazmul0087 and primary data/data from the hospital, using convolutional neural network (CNN) models, namely augmentation, GANs, original, and SMOTE by training and testing. The results of the accuracy value of the training model (dataset nazmul0087) in the detection of kidney cysts, stones, tumors, and normal. The results of augmentation value are 99.93%, GANs 100%, original 100%, and SMOTE 99.93%. In the results of the training model, a very high accuracy value is obtained, with perfect results. The testing model's accuracy value in detecting kidney cysts, stones, tumors, and normal kidney tissue in the original dataset and hospital data. The results of augmentation value are 11.48%, GANs 17.96%, original 21.76%, and SMOTE 20.41%. In the results of the training model, the highest accuracy value is obtained in the original model. For the testing model to automatically diagnose kidney illness and obtain a high accuracy value, which can enhance patient outcomes and save health care costs, we advise using it in conjunction with the original model.
- 1 February 2025
- Published by Institute of Advanced Engineering and Science in IAES International Journal of Artificial Intelligence (IJ-AI)
- Vol. 14 (1) , 602-613
Open AccessUtilization of convolutional neural network in image interpretation techniques for detecting kidney disease
This research is conducted with deep learning for kidney stone disease detection including cysts, stones, normal, and tumors using axial computerized tomography (CT) scan images. The author uses augmentation, generative adversarial networks (GANs), original, and synthetic minority over-sampling technique (SMOTE) to classify kidney disease (cyst, stone, normal, and tumor). This study uses the public dataset nazmul0087 and primary data/data from the hospital, using convolutional neural network (CNN) models, namely augmentation, GANs, original, and SMOTE by training and testing. The results of the accuracy value of the training model (dataset nazmul0087) in the detection of kidney cysts, stones, tumors, and normal. The results of augmentation value are 99.93%, GANs 100%, original 100%, and SMOTE 99.93%. In the results of the training model, a very high accuracy value is obtained, with perfect results. The testing model's accuracy value in detecting kidney cysts, stones, tumors, and normal kidney tissue in the original dataset and hospital data. The results of augmentation value are 11.48%, GANs 17.96%, original 21.76%, and SMOTE 20.41%. In the results of the training model, the highest accuracy value is obtained in the original model. For the testing model to automatically diagnose kidney illness and obtain a high accuracy value, which can enhance patient outcomes and save health care costs, we advise using it in conjunction with the original model.
- Open Access
- 1 December 2024
- Published by Institute of Advanced Engineering and Science in International Journal of Advances in Applied Sciences
- Vol. 13 (4) , 1009-1018
Enhanced U-Net models with encoder and augmentation for phytoplankton segmentation
This study comprehensively analyzes U-Net models for semantic segmentation in phytoplankton image recognition, leveraging encoders such as EfficientNet-B5, MobileNetV2, ResNet50, and ResNeXt50 and employing the Adam optimizer. The research highlights the U-Net MobileNetV2 model with optical distortion, which achieves notable test scores with 93.69% Dice, 88.14% intersection over union (IoU), 99.89% Precision, and 100% Recall, underscoring the efficacy of the applied augmentation strategies, including geometric and distortion transforms, and color and blur techniques. The U-Net ResNet50 model with mix transform consistently demonstrates high accuracy in critical metrics, outperforming others, while EfficientNet-B5 with blur suggests increased model sensitivity with improved recall. These results underscore the crucial role of encoder-augmentation synergy in model performance. Training and testing times across models have remained under 250 seconds, reflecting methodological efficiency. Overall, these results demonstrate the model's excellent performance for the semantic segmentation task.
- 1 December 2024
- Published by Institute of Advanced Engineering and Science in International Journal of Advances in Applied Sciences
- Vol. 13 (4) , 1009-1018
Open AccessEnhanced U-Net models with encoder and augmentation for phytoplankton segmentation
This study comprehensively analyzes U-Net models for semantic segmentation in phytoplankton image recognition, leveraging encoders such as EfficientNet-B5, MobileNetV2, ResNet50, and ResNeXt50 and employing the Adam optimizer. The research highlights the U-Net MobileNetV2 model with optical distortion, which achieves notable test scores with 93.69% Dice, 88.14% intersection over union (IoU), 99.89% Precision, and 100% Recall, underscoring the efficacy of the applied augmentation strategies, including geometric and distortion transforms, and color and blur techniques. The U-Net ResNet50 model with mix transform consistently demonstrates high accuracy in critical metrics, outperforming others, while EfficientNet-B5 with blur suggests increased model sensitivity with improved recall. These results underscore the crucial role of encoder-augmentation synergy in model performance. Training and testing times across models have remained under 250 seconds, reflecting methodological efficiency. Overall, these results demonstrate the model's excellent performance for the semantic segmentation task.