Wyniki wyszukiwania - BazTech

1

Textual explanations for image classification using multimodal LLM

Leś Tomasz, Sawicki Bartosz

Przegląd Elektrotechniczny

|

2025

|

R. 101, nr 4

79--83

EN

In this study, we assess the multimodal capabilities of GPT-4o, focusing on its application to image classification with textual justifications. A series of experiments were conducted, including the recognition of geometric shapes, color differentiation, and melanoma detection using the ISIC skin lesion database. The results indicate that GPT-4o performs comparably to human-level understanding in shape and color recognition, particularly when provided with well-structured prompts. In the medical domain, the model achieved high accuracy in identifying melanoma and nevus lesions based on ABCD criteria. Furthermore, the ability of GPT-4o to provide detailed textual explanations for its decisions enhanced the confidence and transparency of its classifications, making it a promising tool for AI-driven diagnostic support in healthcare.

PL

W niniejszym badaniu oceniono możliwości multimodalne modelu GPT-4o, koncentrując się na jego zastosowaniu w klasyfikacji obrazów z tekstowymi uzasadnieniami. Przeprowadzono serię eksperymentów, w tym rozpoznawanie kształtów geometrycznych, różnicowanie kolorów oraz wykrywanie czerniaka przy użyciu bazy danych zmian skórnych ISIC. Wyniki wskazują, że GPT-4o działa na poziomie zbliżonym do ludzkiego w zakresie rozpoznawania kształtów i kolorów, szczególnie gdy otrzymuje dobrze zdefiniowane polecenia. W dziedzinie medycyny model osiągnął wysoką dokładność w identyfikacji zmian czerniakowych i znamion na podstawie kryteriów ABCD. Ponadto, zdolność GPT-4o do generowania szczegółowych uzasadnień tekstowych dla swoich decyzji zwiększyła zaufanie i przejrzystość jego klasyfikacji, co czyni go obiecującym narzędziem wspierającym diagnostykę opartą na AI w opiece zdrowotnej.

2

Adaptive separation fusion: a novel downsampling approach in CNNS

Ji Xia, Chang Jinglong, Ji Yapeng

Journal of Artificial Intelligence and Soft Computing Research

|

2025

|

Vol. 15, No. 2

197--210

EN

Almost all computer vision tasks rely on convolutional neural networks and transformers, both of which require extensive computations. With the increasingly large size of images, it becomes challenging to input these images directly. Therefore, in typical cases, we downsample the images to a reasonable size before proceeding with subsequent tasks. However, the downsampling process inevitably discards some fine-grained information, leading to network performance degradation. Existing methods, such as strided convolution and various pooling techniques, struggle to address this issue effectively. To overcome this limitation, we propose a generalized downsampling module, Adaptive Separation Fusion Downsampling (ASFD). ASFD adaptively captures intra- and inter-region attentional relationships and preserves feature representations lost during downsampling through fusion. We validate ASFD on representative computer vision tasks, including object detection and image classification. Specifically, we incorporated ASFD into the YOLOv7 object detection model and several classification models. Experiments demonstrate that the modified YOLOv7 architecture surpasses state-of-the-art models in object detection, particularly excelling in small object detection. Additionally, our method outperforms commonly used downsampling techniques in classification tasks. Furthermore, ASFD functions as a plug-and-play module compatible with various network architectures.

3

Exploring data preparation strategies: A comparative analysis of vision transformer and ConvNeXT architectures in breast cancer histopathology classification

Kaczmarek Mikołaj, Kowal Marek, Korbicz Józef

International Journal of Applied Mathematics and Computer Science

|

2025

|

Vol. 35, no. 2

329--339

EN

Breast cancer remains a major global health challenge and the accurate classification of histopathological samples into benign and malignant categories is critical for effective diagnosis and treatment planning. This study offers a comparative analysis of two state-of-the-art deep learning architectures, Vision Transformer (ViT) and ConvNeXT for breast cancer histopathology image classification, focusing on the impact of data preparation strategies. Using the BreakHis benchmark dataset, we investigated six distinct preprocessing approaches, including image resizing, patch-based techniques, and cellular content filtering, applied across four magnification levels (40×, 100×, 200×, and 400×). Both models were fine-tuned and evaluated using multiple performance metrics: accuracy, precision, recall, F1 score, and area under the receiver operating characteristic curve (AUC). The results highlight the critical influence of data preparation on model performance. ViT achieved its highest accuracy of 95.6% and an F1 score of 96.8% at 40× magnification with randomly generated patches. ConvNeXT demonstrated strong robustness across scenarios, attaining a precision of 98.5% at 100× magnification using non-overlapping patches. These findings emphasize the importance of customized data preprocessing and informed model selection in improving diagnostic accuracy. Optimizing both architectural design and data handling is essential to enhancing the reliability of automated histopathological analysis and supporting clinical decision-making.

4

Manual versus Digital Classification of UAV Images in Oak Phenological Studies

Będkowski Krzysztof

Geomatics and Environmental Engineering

|

2025

|

Vol. 19, no. 5

5--22

EN

This research concerns the phenological phenomenon of the autumn discolorations of sessile oak leaves as the trees prepare for winter dormancy. Sessile oak trees were categorized into five classes according to the general colors of their crowns: from green to brown. Low-altitude UAV-acquired images from the visible B, G, and R bands were used, compared, and evaluated against the results of several classification methods: those that were carried out in the field, visually based on orthomosaic observations, and four variants of digital classification. The analysis showed that those methods that were based on observer assessments were highly subjective. At the same time, there was also the problem of the reference data to which the results of the individual methods could be referred. It was expected that the analyzed phenomenon of tree-crown discoloration would be better visible in aerial photographs than in field observations; However, visual color classifications using orthomosaics can be too subjective (as has been shown). It is recommended to use supervised digital classification with a careful selection of reference (training) objects. To switch from pixel-classification results to individual tree classifications, a novel approach was adopted in which the class value that was most abundant within the images of each canopy (determined by the supervised classification method selected) could be used. Among the supervised digital-classification methods that were applied, the results that were closest to the classification performed in the field were obtained by using the ML and Fisher algorithms (followed by kNN).

5

Regional constraint consistency contrastive learning for automatic detection of urinary sediment in microscopic images

Li Fufei, Li Chen, Song Ge, Su Lianzheng, Wang Shian, Fu Qiuyue, Nie Yongqi, Wang Peng

Biocybernetics and Biomedical Engineering

|

2025

|

Vol. 45, iss. 1

74--89

EN

Diagnosing renal and urinary system illnesses usually entails analysing the sediment found in urine. The components in microscopic urine images are diverse and show high similarity, with low contrast due to noise, impeding the progress of automated urine analysis. In order to tackle this difficulty, we propose a region-constrained consistency contrastive learning approach for automated urine analysis. In the first stage, we tackle the complex overlap phenomena in microscopic urine images by innovating the Urine Sediment Paste (US-Paste) positive sample construction method based on supervised contrastive learning. This method uses label information to apply regional constraints and improves the performance of out-of-distribution detection. We also rebuilt the Global Guidance Module (GG Module) and the Enhanced Supervision Module (ES Module). The former improves contrast in urine sediment images by restoring important image details guided by an encoder-decoder structure, while the latter achieves strong feature consistency by combining the most pertinent feature responses from four sets of attention feature maps, which are further mapped via a projection network. In the second phase, we enhance the representations acquired in the initial phase by incorporating a linear classification layer. Our region-constrained consistency contrastive learning algorithm attained an average classification accuracy of 98.30%, precision of 98.33%, recall of 98.30%, and F1-score of 98.30% on the private dataset. Furthermore, in the public urine sediment dataset, the approach achieved an average classification accuracy of 96.19%, precision of 95.79%, recall of 96.19%, and F1-score of 95.94%. The public chromosomal dataset yielded an average classification accuracy of 95.46%, precision of 94.84%, recall of 95.47%, and F1-score of 95.15%. Our methodology surpasses the most advanced methods and demonstrates exceptional performance in urine analysis. This showcases the efficiency of our label-based regional limitations, the outstanding out-of-distribution detection performance of US-Paste, and the robust feature consistency achieved by the Guided Supervision Encoder (GS Encoder). This substantially enhances diagnostic efficiency for clinicians and significantly advances the progress of automated urine sediment analysis.

6

Artificial intelligence models for wound infection recognition and their comparison with human results

Foltynski Piotr, Kruszewska Karolina, Krakowiecki Arkadiusz, Czarkowska-Paczek Bozena, Ladyzynski Piotr

Biocybernetics and Biomedical Engineering

|

2025

|

Vol. 45, iss. 3

572--579

EN

Recognizing an infected wound based solely on a photograph can be a challenge and the aim of this work was to develop a machine learning model that would enable that. We selected 899 wound photographs taken at PODOS Wound Care Clinic (Warsaw, Poland). There were 445 photographs showing uninfected wounds, whereas 454 photographs showed infected wounds with positive microbiological test and antibiotic treatment. A test set was created by randomly selecting 82 photographs representing 42 uninfected and 40 infected wounds. From the remaining photographs, 154 were randomly selected for the validation set, and the remaining 663 formed the training set. Initially we used five pretrained YOLO models from generation 8 and five from generation 11. The 8th generation models performed better than 11th generation models and were then compared with the results of 6 experts and 6 nursing students. The post-hoc analysis revealed that AI models outperformed both specialists and students in terms of mean averaged precision (mAP), accuracy and F1 score, while the results of specialists and students did not differ significantly. For specialists, the medians of mAP, F1 score, and accuracy were 74.1 %, 76.4 %, and 74.4 %, respectively. For Students the medians were 68.4 %, 59.4 %, and 67.7 %, respectively; and for AI models the medians were 92.7 %, 92.9 %, and 92.7 %, respectively. The highest accuracy of 95.1 % of YOLOv8n model was significantly higher than the best specialist’s result of 84.1 %. These results suggest that artificial intelligence can significantly help caregivers recognize wound infection, so they can take appropriate action more quickly.

7

Analysis on grading of lung nodule images with segmentation using u-net and classification with Convolutional Neural Network Fish Swarm Optimization

Sudha R., Uma Maheswari K.M.

Biocybernetics and Biomedical Engineering

|

2025

|

Vol. 45, iss. 1

90--104

EN

Lung malignant tumors are abnormal growths of cells in the lungs that have the potential to invade nearby tissues and spread to other parts of the body. Early detection of these malignant lung tumors is crucial to avoid complications and improve patient outcomes. However, manual processing consumes time and is a tedious process. This might result in poor estimation on cancer-prognosis, leading the patients into a higher risk of mortality. Many existing literatures have detected the malignant tumors, yet, found certain difficulties with the identification of size, appearance and spread of cancerous-cells in lung region to determine how far it has been occupied. Hence, the present study aims to overcome the existing complications through Deep Learning based Swarm Intelligence Algorithms. Implementation of the proposed work is involved with three stages such as preprocessing, segmentation and classification. Besides, CT scan possess the capability for giving a comprehensive view than X-rays. Data are collected from LIDC-IDRI (Lung Image Database Consortium-Image Database Resource Initiative) with lung CT-images and accomplishes pre-processing by removing noise efficiently using wiener filter. Further, changes in soft tissues of lungs are identified and segmented in the subsequent phase using U-Net and finally classification is performed using CFSO (Convolutional Neural Network Fish Swarm Optimization) to overcome the slight chance of misclassification error as proposed CFSO can lead to more efficient computational processes since FSO algorithms are designed to minimize computational costs while maximizing performance through their metaheuristic nature. This efficiency is particularly beneficial when dealing with large datasets typical in medical imaging, allowing faster processing times without sacrificing accuracy. Hence, amalgamation of CFSO can reduce the number of features, thus speeding up training and inference times. Through the performance assessment, IoU (Intersection over Union) value attained through the analysis is found to be 0.7822. Further, accuracy obtained by the proposed model is 97.80%, recall is 98.49%, precision is 96.8% and F1-score is 97.32%. Findings of the study exhibits the purposefulness of the study in clinical settings by potentially reducing false negatives in lung cancer screening, ultimately improving patient survival rates through earlier detection and treatment.

8

Pretrained models for classification of dementia degree and treatment of Alzheimer’s disease

Benbakreti Soumia, Benbakreti Samir, Benyahia Kadda, Khobzaoui Abdelkader

Przegląd Elektrotechniczny

|

2024

|

R. 100, nr 5

135--138

EN

The early diagnosis of Alzheimer’s disease poses a significant challenge in the health sector, and the integration of deep learning and artificial intelligence (AI) holds promising potential for enhancing early detection through the classification of dementia levels, enabling more effective disease treatment. Deep neural networks have the capacity to autonomously learn and identify discriminative characteristics associated with this pathology. In this study, three pre-trained CNN-based models are employed to classfify MRI images of Alzheimer’s patients, with ResNet18 yielding excellent results and achieving an accuracy rate of 97.3%.

PL

Wczesna diagnoza choroby Alzheimera stanowi powazne wyzwanie w sektorze zdrowia, a integracja głębokiego uczenia się i sztucznej inteligencji (AI) niesie obiecujący potencjał w zakresie poprawy wczesnego wykrywania poprzez klasyfikację poziomów demencji, umozliwiając skuteczniejsze leczenie chorób. Głębokie sieci neuronowe mają zdolność autonomicznego uczenia się i identyfikowania cech dyskryminacyjnych związanych z tą patologią. W tym badaniu do klasyfikacji obrazów MRI pacjentów z chorobą Alzheimera wykorzystano trzy wstępnie wyszkolone modele oparte na CNN, przy czym ResNet18 zapewnia doskonałe wyniki i osiąga współczynnik dokładności wynoszący 97,3%

9

Image classification of leukemia cancer using Wavelet deep neural network

Habchi Yassine, Bouddou Riyadh, Aimer Ameur Fethi

Przegląd Elektrotechniczny

|

2024

|

R. 100, nr 3

238--243

EN

Classification of blood cell images, through color and morphological features, is essential for medical diagnostic processes. This paper proposes an efficient method using LeGall5/3 wavelet transform (LeGall5/3WT) based on Convolutional Neural Network (CNN) for leukemia cancer image classification. The proposed algorithm is applied on 108leukemia images, including 49 blast cell images and 59 healthy cell images. All these images are obtained from the acute lymphoblastic leukemia image database for image processing (ALL-IDB). The data augmentation technique provided 7776 images, including3528 blast cell images and 4248 healthy cells. LeGall5/3WT feature extraction results are used as inputs to the CNN for leukemia cancer classification. The network system architecture contains three convolutions, three aggregate layers, a fully connected layer, a Soft Max layer, and an output layer with two classes. The proposed algorithm achieves accurate results (accuracy of 100%, sensitivity of 100%, specificity of 100%) for ALL-LDB1 database.

PL

Klasyfikacja obrazów krwinek pod kątem cech kolorystycznych i morfologicznych jest niezbędna w procesach diagnostyki medycznej. W artykule zaproponowano wydajną metodę wykorzystującą transformatę falkową LeGall5/3 (LeGall5/3WT) opartą na konwolucyjnej sieci neuronowej (CNN) do klasyfikacji obrazów raka białaczki. Proponowany algorytm jest stosowany na 108 obrazach białaczki, w tym 49 obrazach komórek blastycznych i 59 obrazach zdrowych komórek. Wszystkie te obrazy uzyskano z bazy danych obrazów ostrej białaczki limfoblastycznej do przetwarzania obrazów (ALL-IDB). Technika powiększania danych dostarczyła 7776 obrazów, w tym 3528 obrazów komórek blastycznych i 4248 zdrowych komórek. Wyniki ekstrakcji cech LeGall5/3WT są wykorzystywane jako dane wejściowe do CNN w celu klasyfikacji raka białaczki. Architektura systemu sieciowego zawiera trzy sploty, trzy warstwy agregatów, warstwę w pełni połączoną, warstwę Soft Max i warstwę wyjściową z dwiema klasami. Zaproponowany algorytm pozwala uzyskać dokładne wyniki (dokładność 100%, czułość 100%, specyficzność 100%) dla bazy danych ALL-LDB1.

10

A MZI-based optical neural network for image classification

Zhang Ye, Zhang Saining, Zhang Danni, Wang Ruiting, Su Yanmei, Yi Junkai, Wang Pengfei, Luo Guangzhen, Zhou Xuliang, Pan Jiaoqing

Optica Applicata

|

2024

|

Vol. 54, nr 1

97--104

EN

In recent years, with the expansion of information, artificial intelligence technology has been developed and used in various fields. Among them, optical neural network provides a new type of special neural network accelerator chip solution, which has the advantages of high speed, high bandwidth, and low power consumption. In this paper, we construct an optical neural network based on Mach–Zehnder interferometer. The experimental results on the image classification of MNIST handwritten digitals show that the optical neural network has high accuracy, fast convergence and good scalability.

11

An evaluation of the capabilities of image-based metal component defect recognition with deep learning techniques

Wójcik Michał P., Pawlikowski Kacper, Madej Łukasz

Computer Methods in Materials Science

|

2024

|

Vol. 24, No. 3

33--40

EN

In the era of Industry 4.0, deploying highly specialised machine learning models trained on unique and often scarce datasets is an attractive solution for advancing automated quality control and minimising production costs. Therefore, the main aim of this research is to evaluate the capabilities of three deep learning models (ResNet-18, ResNet-50 and SE-ResNeXt-101 (32 × 4d)) in the identification of surface defects in forged products. Leveraging advanced photography techniques, including studio lighting and a shadowless box, high-quality images of complex product surfaces were acquired for the training data set. Given the relatively small size of the image dataset, aggressive data augmentation techniques were introduced during the training and evaluation process to ensure robust model generalisation ability. The results obtained demonstrate the significant impact of data augmentation on model performance, highlighting its importance in training and evaluating deep learning models with limited data. This research also emphasises the need for innovative data pre-processing strategies in an efficient and robust machine learning model delivery to the industrial environment.

12

EO-CNN: Equilibrium Optimization-Based hyperparameter tuning for enhanced pneumonia and COVID-19 detection using AlexNet and DarkNet19

Kiziloluk Soner, Sert Eser, Hammad Mohamed, Tadeusiewicz Ryszard, Pławiak Paweł

Biocybernetics and Biomedical Engineering

|

2024

|

Vol. 44, no. 3

635--650

EN

Convolutional neural networks (CNN) have been increasingly popular in image categorization in recent years. Hyperparameter optimization is a critical stage in enhancing the effectiveness of CNNs and achieving better results. Properly tuning hyperparameters allows the model to exhibit improved performance and facilitates faster learning. Misconfigured hyperparameters can prolong the training time or lead to the model not learning at all. Manually tuning hyperparameters is a time-consuming and challenging process. Automatically adjusting hyperparameters helps save time and resources. This study aims to propose an approach that shows higher classification performance than unoptimized convolutional neural network models, even at low epoch values, by automatically optimizing the hyperparameters of AlexNet and DarkNet19 with equilibrium optimization, the newest metaheuristic algorithm. In this respect, the proposed approach optimizes the number and size of filters in the first five convolutional layers in AlexNet and DarkNet19 using an equilibrium optimization algorithm. To evaluate the efficacy of the suggested method, experimental analyses were conducted on the pneumonia and COVID-19 datasets. An important advantage of this approach is its ability to accurately classify medical images. The testing process suggests that utilizing the proposed approach to optimize hyperparameters for AlexNet and DarkNet19 led to a 7% and 4.07% improvement, respectively, in image classification accuracy compared to non-optimized versions of the same networks. Furthermore, the approach displayed superior classification performance even in a few epochs compared to AlexNet, ShuffleNet, DarkNet19, GoogleNet, MobileNet-V2, VGG-16, VGG-19, ResNet18, and Inceptionv3. As a result, automatic tuning of the hyperparameters of AlexNet and DarkNet-19 with EO enabled the performance of these two models to increase significantly.

13

The Use of Artificial Intelligence for Quality Assessment of Refill Friction Stir Spot Welded Thin Joints

Kubit Andrzej, Kłosowski Grzegorz, Berezowski Wojciech

Advances in Science and Technology. Research Journal

|

2024

|

Vol. 18, no. 3

45--57

EN

This paper presents a machine learning and image segmentation based advanced quality assessment technique for thin Refill Friction Stir Spot Welded (RFSSW) joints. In particular, the research focuses on developing a predictive support vector machines (SVM) model. The purpose of this model is to facilitate the selection of RFSSW process parameters in order to increase the shear load capacity of joints. In addition, an improved weld quality assessment algorithm based on optical analysis was developed. The research methodology includes specimen preparation stages, mechanical tests, and algorithmic analysis, culminating in a machine learning model trained on experimental data. The results demonstrate the effectiveness of the model in selecting welding process parameters and assessing weld quality, offering significant improvements compared to standard techniques. This research not only proposes a novel approach to optimizing welding parameters but also facilitates automatic quality assessment, potentially revolutionizing and spreading the application of the RFSSW technique in various industries

14

Identifying selected diseases of leaves using deep learning and transfer learning models

Mimi Afsana, Zohura Sayeda Fatema Tuj, Ibrahim Muhammad, Haque Riddho Ridwanul, Farrok Omar, Jabid Taskeed, Ali Md Sawkat

Machine Graphics and Vision

|

2023

|

Vol. 32, No. 1

55--71

EN

Leaf diseases may harm plants in different ways, often causing reduced productivity and, at times, lethal consequences. Detecting such diseases in a timely manner can help plant owners take effective remedial measures. Deficiencies of vital elements such as nitrogen, microbial infections and other similar disorders can often have visible effects, such as the yellowing of leaves in Catharanthus roseus (bright eyes) and scorched leaves in Fragaria ×ananassa (strawberry) plants. In this work, we explore approaches to use computer vision techniques to help plant owners identify such leaf disorders in their plants automatically and conveniently. This research designs three machine learning systems, namely a vanilla CNN model, a CNN-SVM hybrid model, and a MobileNetV2-based transfer learning model that detect yellowed and scorched leaves in Catharanthus roseus and strawberry plants, respectively, using images captured by mobile phones. In our experiments, the models yield a very promising accuracy on a dataset having around 4000 images. Of the three models, the transfer learning-based one demonstrates the highest accuracy (97.35% on test set) in our experiments. Furthermore, an Android application is developed that uses this model to allow end-users to conveniently monitor the condition of their plants in real time.

15

Additional Look into GAN-based Augmentation for Deep Learning COVID-19 Image Classification

Fedoruk Oleksandr, Klimaszewski Konrad, Ogonowski Aleksander, Kruk Michał

Machine Graphics and Vision

|

2023

|

Vol. 32, No. 3/4

107--124

EN

Data augmentation is a popular approach to overcome the insufficiency of training data for medical imaging. Classical augmentation is based on modification (rotations, shears, brightness changes, etc.) of the images from the original dataset. Another possible approach is the usage of Generative Adversarial Networks (GAN). This work is a continuation of the previous research where we trained StyleGAN2-ADA by Nvidia on the limited COVID-19 chest X-ray image dataset. In this paper, we study the dependence of the GAN-based augmentation performance on dataset size with a focus on small samples. Two datasets are considered, one with 1000 images per class (4000 images in total) and the second with 500 images per class (2000 images in total). We train StyleGAN2-ADA with both sets and then, after validating the quality of generated images, we use trained GANs as one of the augmentations approaches in multi-class classification problems. We compare the quality of the GAN-based augmentation approach to two different approaches (classical augmentation and no augmentation at all) by employing transfer learning-based classification of COVID-19 chest X-ray images. The results are quantified using different classification quality metrics and compared to the results from the previous article and literature. The GAN-based augmentation approach is found to be comparable with classical augmentation in the case of medium and large datasets but underperforms in the case of smaller datasets. The correlation between the size of the original dataset and the quality of classification is visible independently from the augmentation approach.

16

Melanoma skin cancer and nevus mole classification using intensity value estimation with convolutional neural network

Ashafuddula N. I. Md., Islam Rafiqul

Computer Science

|

2023

|

T. 24 (3)

277--296

EN

Melanoma skin cancer is one of the most dangerous and life-threatening cancer. Exposure to ultraviolet rays may damage the skin cell's DNA, which can causes melanoma skin cancer. However, detecting and classifying melanoma and nevus moles at their immature stages is difficult. In this work, an automatic deep-learning system has been developed based on intensity value estimation with a convolutional neural network model (CNN) for detecting and classifying melanoma and nevus moles more accurately. Since intensity levels are the most distinctive features for identifying objects or regions of interest, high-intensity pixel values have been selected from extracted lesion images. Incorporating those high-intensity features into CNN improves the overall performance of the proposed model than the state-of-the-art methods for detecting melanoma skin cancer. To evaluate the system, we used five-fold cross-validation. The experimental results showed that superior percentages of accuracy (92.58%), sensitivity (93.76%), specificity (91.56%), and precision (90.68%) were achieved.

17

Attention-guided multiple instance learning for COPD identification: To combine the intensity and morphology

Wu Yanan, Qi Shouliang, Feng Jie, Chang Runsheng, Pang Haowen, Hou Jie, Li Mengqi, Wang Yingxi, Xia Shuyue, Qian Wei

Biocybernetics and Biomedical Engineering

|

2023

|

Vol. 43, no. 3

568--585

EN

Chronic obstructive pulmonary disease (COPD) is a complex and multi-component respiratory disease. Computed tomography (CT) images can characterize lesions in COPD patients, but the image intensity and morphology of lung components have not been fully exploited. Two datasets (Dataset 1 and 2) comprising a total of 561 subjects were obtained from two centers. A multiple instance learning (MIL) method is proposed for COPD identification. First, randomly selected slices (instances) from CT scans and multi-view 2D snapshots of the 3D airway tree and lung field extracted from CT images are acquired. Then, three attention-guided MIL models (slice-CT, snapshot-airway, and snapshot-lung-field models) are trained. In these models, a deep convolution neural network (CNN) is utilized for feature extraction. Finally, the outputs of the above three MIL models are combined using logistic regression to produce the final prediction. For Dataset 1, the accuracy of the slice-CT MIL model with 20 instances was 88.1%. The backbone of VGG-16 outperformed Alexnet, Resnet18, Resnet26, and Mobilenet_v2 in feature extraction. The snapshotairway and snapshot-lung-field MIL models achieved accuracies of 89.4% and 90.0%, respectively. After the three models were combined, the accuracy reached 95.8%. The proposed model outperformed several state-of-the-art methods and afforded an accuracy of 83.1% for the external dataset (Dataset 2). The proposed weakly supervised MIL method is feasible for COPD identification. The effective CNN module and attention-guided MIL pooling module contribute to performance enhancement. The morphology information of the airway and lung field is beneficial for identifying COPD.

18

Rotation-Gamma Correction augmentation on CNN-Dense block for soil image classification

Maiyanti Sri Indra, Desiani Anita, Lamin Syafrina, Puspitashati, Arhami Muhammad, Gofar Nuni, Cahyana Destika

Applied Computer Science

|

2023

|

Vol. 19, no. 3

96--115

EN

Soil is a solid particle that covers the surface of the earth. Soil can be classified based on its color because the color indicates the nature and condition of the soil. CNN works well for image classification, but it requires large amounts of data. Augmentation is a technique to increase the amount of training data with various transformation techniques to the existing data. Rotation and Gamma Correction can be used simply as an augmentation technique and can reproduce an image with as many image variations as desired from the original image. CNN architecture has a convolution layer and Dense block has dense layers. The addition of Dense blocks to CNN aims to overcome underfitting and overfitting problems. This study proposes a combination of Augmentation and classification. In augmentation, a combination of rotation and Gamma correction techniques is used to reproduce image data. The CNN-Dense block is applied for classification. The soil image classification is grouped based on 5 labels black soil, cinder soil, laterite soil, peat soil, and yellow soil. The performances of the proposed method provide excellent results, where accuracy, precision, recall, and F1-Score performances are above 90%. It can be concluded that the combination of rotation and Gamma Correction as augmentation techniques and CNN-Dense blocks is powerful for use in soil image classification.

19

Analiza możliwości automatycznej detekcji obiektów topograficznych na zdjęciach lotniczych i satelitarnych VHR

Pluto-Kossakowska Joanna, Kamiński Michał

Teledetekcja Środowiska

|

2022

|

T. 62

5--15

PL

W artykule podjęto temat uczenia maszynowego w rozpoznawaniu obiektów topograficznych na zdjęciach lotniczych i satelitarnych VHR ze szczególnym uwzględnieniem Bazy Danych Obiektów Topograficznych BDOT10k. Celem prac badawczych było przetestowanie trzech algorytmów klasyfikacji nadzorowanej do automatycznej detekcji wybranych klas obiektów topograficznych, m.in.: budynków, betonowych oraz asfaltowych elementów szarej infrastruktury (drogi, chodniki, place), wód powierzchniowych, lasów, terenów zadrzewionych i zakrzewionych, terenów o niskiej roślinności oraz gleby odkrytej (grunty nieużytkowane, wyrobiska). Przeanalizowano trzy powszechnie stosowane klasyfikatory: Maximum Likelihood, Support Vector Machine oraz Random Trees pod kątem różnych parametrów wejściowych. Wynikiem przeprowadzonych badań jest ocena ich skuteczności w detekcji poszczególnych klas oraz ocena przydatności wyników klasyfikacji do aktualizacji bazy danych BDOT10k. Badania zostały przeprowadzone dla zdjęcia satelitarnego WorldView-2 o rozdzielczości przestrzennej 0,46 m oraz ortofotomapy ze zdjęć lotniczych o dokładności przestrzennej 0,08 m. Wyniki badań wskazują na to, że wykorzystanie różnych klasyfikatorów uczenia maszynowego oraz danych źródłowych wpływa nieznacznie na wynik klasyfikacji. Ogólne statystyki dokładności wskazują, że całościowo klasyfikacja z wykorzystaniem zdjęć satelitarnych dała nieco lepsze rezultaty o kilka punktów procentowych w granicach 76-81%, a dla zdjęć lotniczych 75-78%. Natomiast dla niektórych klas miara statystyczna F1 przekracza wartość 0,9. Testowane algorytmy uczenia maszynowego dają bardzo dobre rezultaty w identyfikacji wybranych obiektów topograficznych, ale nie można jeszcze mówić o bezpośredniej aktualizacji BDOT10k.

EN

The article deals with the topic of machine learning (ML) in the recognition of topographic objects in aerial and satellite VHR image, with particular emphasis on the Topographic Objects Database (BDOT10k). The aim of the research work was to test three supervised classification algorithms for automatic detection of selected classes of topographic objects, including: buildings, concrete and asphalt elements of grey infrastructure (roads, pavements, squares), surface waters, forests, wooded and bushy areas, areas with low vegetation and uncovered soil (unused lands or excavations). Three commonly used classifiers were analysed: Maximum Likelihood, Support Vector Machine and Random Trees for different input parameters. The result of the research is the assessment of their effectiveness in the detection of individual classes and the assessment of the suitability of the classification results for updating the BDOT10k database. The research was carried out for the WorldView-2 satellite image with a spatial resolution of 0.46 m and orthophotos from aerial images with a spatial resolution of 0.08 m. The research results indicate that the use of different ML classifiers and source data slightly affects the classification result. Overall accuracy statistics show that the classification using satellite images gave slightly better results by a few percentage points in the range from 76% to 81%, and for aerial photos from 75% to 78%. However, for some classes the statistical measure F1 exceeds 0.9 value. The tested ML algorithms give very good results in identifying selected topographic objects, but it is not yet possible to directly update topographical database.

20

Fourier transform-based data augmentation in deep learning for diabetic foot thermograph classification

Anaya-Isaza Andrés, Zequera-Diaz Martha

Biocybernetics and Biomedical Engineering

|

2022

|

Vol. 42, no. 2

437--452

EN

Diabetes Mellitus (DM) belongs to the ten diseases group with the highest mortality rate globally, with an estimated 578 million cases by 2030, according to the World Health Organization (WHO). The disease manifests itself through different disorders, where vasculopathy shows a chronic relationship with diabetic ulceration events in distal extremities, being temperature a biomarker that can quantify the risk scale. According to the above, an analysis is performed with standing thermography images, finding temperature patterns that do not follow a particular distribution in patients with DM. Therefore, the modern medical literature has taken the use of Computer-Aided Diagnosis (CAD) systems as a plausible option to increase medical analysis capabilities. In this sense, we proposed to study three state-of-the-art deep learning (DL) architectures, experimenting with convolutional, residual, and attention (Transformers) approaches to classify subjects with DM from diabetic foot thermography images. The models were trained under three conditions of data augmentation. A novel method based on modifying the images through the change of the amplitude in the Fourier Transform is proposed, being the first work to perform such synergy in the characterization of risk in ulcers through thermographies. The results show that the proposed method allowed reaching the highest values, reaching a perfect classification through the convolutional neural network ResNet50v2, promising for limited data sets in thermal pattern classification problems.