KMITL

Permanent URI for this communityhttps://dspace.kmitl.ac.th/handle/123456789/1

Browse

Search Results

Now showing 1 - 10 of 29

Some of the metrics are blocked by your
consent settings
Comparative performance of deep learning models and non-dermatologists in diagnosing psoriasis, dermatophytosis, and eczema
(2026-12-01)
Yodrabum, Nutcha
;
Wongpraparut, Chanisada
;
Titijaroonroj, Taravichet
;
Chularojanamontri, Leena
;
Bunyaratavej, Sumanas
Accurately differentiating scaly erythematous rashes among psoriasis, eczema, and dermatophytosis remains a clinical challenge, particularly for non-dermatologists. This study aimed to develop and evaluate deep learning models using macroscopic clinical images to classify these conditions and compare their performance with that of non-specialists. A total of 2940 images were sourced from public datasets, the Siriraj Dermatology databank, and newly collected images from Thai participants. Among sixteen evaluated models, the Swin demonstrated the best performance and interpretability. Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations confirmed that the model focused on clinically relevant lesion features. Most importantly, in a pilot comparison, the Swin outperformed non-specialists in diagnostic accuracy. However, given the limited sample size of 30 images and 30 evaluators, these results should be interpreted as exploratory. Future studies with larger datasets and diverse clinician cohorts are warranted to confirm these findings and to support clinical integration.
Some of the metrics are blocked by your
consent settings
Deep learning-based object detection of restorative dental instruments with potential implications for workflow automation and infection control in dental supply units
(2026-12-01)
Poomrittigul, Suvit
;
Mittong, Sirawit
;
Thanathornwong, Bhornsawan
;
Suebnukarn, Siriwan
This study presents a proof-of-concept deep learning approach for automated detection and classification of restorative dental instruments on standardized trays, aiming to support workflow automation and infection control in dental supply units. A dataset comprising 1,000 images and 14,000 annotated instances of restorative dental instruments across 14 categories was developed. The YOLOv8 model was trained and evaluated on this dataset using standard object detection metrics, including precision, recall, and mean average precision at IoU thresholds 0.5 (mAP@0.5) and 0.5:0.95 (mAP@[0.5:0.95]). To assess model advancement, YOLOv8 performance was compared against its predecessors, YOLOv5, YOLOv6, and YOLOv7, under identical experimental settings. A session-level data split was implemented as the primary evaluation to minimize data leakage and provide a realistic estimate of generalization across unseen tray configurations. The YOLOv8 model achieved highest mean average precision mAP@0.5 of 95.9% and mAP@[0.5:0.95] of 80.9%, demonstrating robust detection capability under both standard and stringent evaluation thresholds. Across instrument categories, YOLOv8 demonstrated precision ranging from 90.3% to 100% and recall from 80.6 to 98.5%. The findings demonstrate the feasibility of using YOLOv8 for automated restorative dental instrument detection as an early-stage tool for improving supply unit efficiency. While results indicate high detection accuracy and robustness, further validation in diverse clinical environments is needed. Future deployment should incorporate human-in-the-loop verification, audit trails, and error escalation mechanisms to ensure safe and accountable AI-assisted workflows.
Some of the metrics are blocked by your
consent settings
Systematic evaluation of one-dimensional-to-two-dimensional near-infrared spectroscopy transformations with deep learning for quantifying coconut sap adulteration
(2026-10-01)
Lapcharoensuk, Ravipat
;
Sitorus, Agustami
Near-infrared (NIR) spectroscopy have limitations when combined with deep learning (DL) algorithms because they rely on low-dimensional datasets. Therefore, we investigated the potential of transforming one-dimensional (1D) NIR spectra into two-dimensional (2D) spectrograms using synchronous and asynchronous techniques and the continuous wavelet transform (CWT) and their effectiveness by integrating with DL for detecting adulteration in coconut sap. NIR spectra (12,500–4000 cm<sup>−1</sup>) were collected from binary mixtures (0%–100%; w /w). The performance of all DL (convolutional neural networks-CNN, AlexNet and ResNet) models was compared with that of partial least squares (PLS). The models were ranked in the mentioned order based on their performances: 2D-CWT ' 2D-asynchronous ' 2D-synchronous ' 1D/2D-PLS. The important features of the best model can be explained and visualized using gradient-weighted-class-activation-mapping. The findings highlight that the 1D-to-2D NIR data transformation combined with DL is a highly robust approach because it addresses the feature representation gap in NIR data and effectively captures the spatial–spectral correlations.
Some of the metrics are blocked by your
consent settings
Huanglongbing (HLB) disease detection using drone imagery and deep learning neural networks for early management of HLB
(2026-08-01)
Kalbande, Vishal Dashrathrao
;
Phanomsophon, Thitima
;
Hongwiangjan, Jeerayut
;
Sripinyowanich Jongyingcharoen, Jiraporn
;
Sirisomboon, Panmanas
In the modern world, machine learning and artificial intelligence have become the foundation of the digital revolution and have a significant role in daily life. It is adopted in various applications, such as object detection, recognition and classification. This research aimed to use deep learning algorithms for the detection of Huanglongbing (HLB)-infected disease, healthy and background patch in citrus orchard. HLB-infected patches need to be identified for appropriate treatments of spraying. However, farmers are unaware of infections on plant leaves and therefore adopt manual disease identification methods. This method results in loss of productivity as the infection spreads throughout the field. However, due to a lack of required facilities, instant identification needs to be improved in many aspects of the agricultural sector. To conduct this research, firstly, a dataset was created that contained drone images of citrus orchards that was categorized into three classes, specifically (i) Healthy, (ii) HLB-infected and (iii) Background. Under this research, 6000 image patches of citrus orchard were collected and categorized 2000 images in each class based on appropriate labels. The next step was to train the deep learning models to identify Healthy, HLB-infected and Background. In this research three models viz. EfficientNetV2B0, DenseNet-121 and ResNet-50 were trained to detect the HLB infection patch in the citrus orchard. The models exhibited exceptional performance, with ResNet-50 achieving the highest overall accuracy of 89%, followed by EfficientNet-V2B0 (87%) and DenseNet-121 (85%). ResNet-50 demonstrated superior balanced performance with macro-average precision of 90%, recall of 89%, and F1-score of 89%. The generated HLB detection map effectively visualized disease distribution patterns, identifying 31% of patches as HLB-infected, 7% as Healthy and 62% as Background across the orchard grid. This research establishes a reliable framework for large-scale, automated citrus disease monitoring that enables precision agriculture interventions, potentially reducing economic losses and optimizing resource allocation in citrus cultivation through early and accurate HLB disease detection.
Some of the metrics are blocked by your
consent settings
High-speed multicontrast dynamic OCT by using deep learning
(2026-03-05)
Liu, Yusong
;
Abd El-Sadek, Ibrahim
;
Morishita, Rion
;
Padungatthakij, Chettanat
;
Furukawa, Atsuko
We proposed a deep learning approach to significantly accelerate multi-contrast dynamic optical coherence tomography (MC-DOCT) imaging. We trained a three-dimensional UNet with a long short-Term memory module to simultaneously generate authentic logarithmic intensity variance (aLIV) and swiftness images from only 4 OCT frames with nonuniform time intervals, instead of conventional 32 frames. This method accurately visualized multiple functional domains in in vitro samples such as breast cancer spheroids, colon cancer spheroids and alveolar organoids, which is consistent with ground truth computed from 32 OCT frames. Our method provides a promising solution for reducing MC-DOCT acquisition time from 1 minute to 6 seconds.
Some of the metrics are blocked by your
consent settings
Enhancing deep learning–based railway inspection via PSO-guided brightness–contrast optimization
(2026-03-01)
Songthai, Maethinee
;
Kaitwanidvilai, Somyot
Reliable operation of electrified railway systems depends critically on the pantograph–catenary system (PCS). Existing inspection practices are largely manual or periodic and remain vulnerable to low illumination, background clutter, and thin structural components, limiting robustness and scalability. Although vision-based deep learning is promising, performance often degrades in low-light and complex scenes, while conventional enhancement (e.g., CLAHE) provides limited and inconsistent improvements. This study proposes an illumination-aware PCS inspection framework that integrates Particle Swarm Optimization (PSO)–guided brightness–contrast optimization with a lightweight YOLO detector. A version benchmark selected YOLOv9t as the backbone, achieving the best overall performance (Precision = 0.947, F1-score = 0.913) compared with YOLOv8n and YOLOv11n. Although YOLOv11n has lower FLOPs, PCS inspection is accuracy-critical because detection errors propagate to event-frequency counting and lateral-deviation assessment; therefore, YOLOv9t was fixed for subsequent experiments. PSO is employed to estimates a single dataset-level global enhancement parameter set, improving robustness while maintaining computational efficiency for early-stage field deployment under limited annotated data and edge hardware constraints. The framework was evaluated on a small yet diverse visible-light dataset spanning day/evening/night conditions and challenging locations (station roofs, bracket regions, and transition sections). Comparative evaluation across three configurations—YOLOv9 baseline, YOLOv9 with CLAHE, and YOLOv9 with PSO-tuned brightness/contrast—achieved detection rates of 90%, 79%, and 99%, respectively, with the largest gains on thin, low-contrast structures (contact and messenger wires). The optimized pipeline further enables reliable estimation of lateral wire deviation in compliance with EN 50367 (≤200 mm), supporting safety-critical inspection in low-light and complex-background conditions.
Some of the metrics are blocked by your
consent settings
Multi-agent deep learning on tensor fields for segmentation of ultrasound images
(2026-03-01)
Sharma, Suman
;
Moodleah, Samart
;
Makhanov, Stanislav S.
Medical image analysis often relies on vector fields (VF), which are fundamental to deterministic models such as Active Contours, Level Set Methods, Phase Portrait Analysis, and artificial agent–based formulations. We experimentally demonstrate that a Deep Learning Neural Network (DLNN) capable of interpreting VF structures can substantially enhance the decision-making capabilities of artificial agents. We introduce a novel hybrid framework that integrates artificial life (AL) agents operating within a VF with a DLNN that guides their behavior. A key innovation of the model is the initialization of AL agents using streamlines derived from the VF orthogonal to the generalized gradient vector flow (GGVF) field. The VF is further transformed into a bi-directional Tensor Field (TF), where the spatial distribution and classification of degenerate points (DPs) serve as critical features. These DPs are leveraged to train AL agents through the DLNN, enabling them to follow meaningful anatomical structures. The framework employs DeepLabV3+ with ResNet50 as the backbone and is trained on 179 benign and 107 malignant breast ultrasound images collected at Thammasat University Hospital (TUH) and annotated by three leading radiologists, in addition to the BUSI and UDIAT datasets. Using 10-fold cross-validation, the proposed method achieves stable and robust performance across three datasets. Mean Dice scores of 94.84±1.63% (TUH), 94.16±1.62% (BUSI), and 93.67±1.51% (UDIAT) are obtained, with corresponding IoU values of 91.19±1.76%, 90.21±1.83% and 89.08±1.70%, demonstrating strong generalization across diverse imaging conditions. Comparative evaluations against state-of-the-art methods confirm the superiority of the proposed model. A video demonstration is available at: https://tinyurl.com/AL-DLNN.
Some of the metrics are blocked by your
consent settings
Enhanced deep learning model for prediction of diabetic mellitus on optical coherence tomography angiography images
(2026-01-01)
Visitsattapongse, Sarinporn
;
Rithcharung, Preeyarat
;
Santiprabhob, Jeerunda
;
Lertbannaphong, Ornsuda
;
Sermsripong, Wasawat
Background: Diabetes mellitus (DM) is a chronic metabolic disease characterized by dysregulated blood glucose. Prolonged DM can lead to diabetic retinopathy (DR), in which retinal capillaries are damaged by sustained hyperglycemia. Optical coherence tomography angiography (OCTA) is a non-invasive imaging modality for visualizing retinal microvasculature and can detect early changes in both DM patients with and without DR. However, it requires expert evaluation, making early detection costly and time-consuming. This study aimed to develop a high-performance deep learning framework that can classify OCTA images into three groups of DM, such as normal, good glycemic control, and poor glycemic control. Methods: OCTA datasets of horizontal B-scans and en face scans from 300 participants aged 8–18 years were analyzed, including normal controls, DM patients with good glycemic control, and DM patients with poor control (HbA1c ≥8%). For each participant, a 3 mm × 3 mm foveal-centered en face image of the deep capillary plexus (DCP) and a horizontal B-scan through the foveal center of the right eye were selected. Several convolutional and transformer-based models were evaluated, with ConvNeXt (a ConvNet for the 2020s) chosen as the baseline for its superior performance. To enhance generalization and convergence, progressive resizing and the Lookahead optimization strategy were applied, while class-wise augmentation was used to balance the training set without altering the test distribution. Results: The baseline ConvNeXt achieved F1 scores of 0.7877 (B-scans) and 0.7424 (en face). After doing enhancement using progressive resizing and Lookahead optimization, performance improved to 0.8319 and 0.8567 (Wilcoxon signed-rank tests, P<0.05). Conclusions: Our proposed method for DM classification from OCTA images provided promising results while ensuring resource efficiency and rapid evaluation. Clinically, accurate classification of DM status is valuable for assessing the risk of DR progression. Thus, it can be served as an assistive tool for clinical decision support in DR management.
Some of the metrics are blocked by your
consent settings
SkinHRNet: A Deep Learning Framework for Non-Contact Heart Rate Estimation and Arterial–Venous Sufficiency Status Classification from Short Skin Videos
(2026-01-01)
Traivinidsreesuk, Chetsadaporn
;
Yodrabum, Nutcha
;
Winaikosol, Kengkart
;
Chaikangwan, Irin
;
Prompattanapakdee, Jiraya
Traditional methods for monitoring flap health in reconstructive surgery are often invasive and rely on subjective assessment. This study addresses clinically motivated monitoring challenges by evaluating two tasks using a dataset of 1,018 short skin videos: average heart rate (HR) estimation under arterial–venous sufficiency conditions and arterial–venous sufficiency status classification across sufficiency and simulated insufficiency conditions. To address these challenges, we propose SkinHRNet, a deep learning–based approach for average HR estimation and arterial–venous sufficiency status classification from short skin videos. This work contributes a short-skin-video framework that combines HR-related signal estimation with arterial–venous sufficiency classification to support the two evaluated tasks under the controlled conditions considered in this study. For average HR estimation under arterial–venous sufficiency conditions, SkinHRNet achieved a mean absolute error (MAE) of 8.66 ± 4.85 BPM. For arterial–venous sufficiency status classification, it achieved an accuracy of 0.969 ± 0.02 across the evaluated sufficiency and simulated insufficiency conditions. These findings indicate that SkinHRNet may serve as an initial research prototype for further investigation of short-video-based non-contact assessment under controlled arterial–venous sufficiency and simulated insufficiency conditions.
Some of the metrics are blocked by your
consent settings
Time Series Forecasting Using Transfer Learning with an Attention-Revamped Transformer: A Case Study of Financial Instruments
(2025-12-01)
Feng, Ling
;
Sinchai, Ananta
Accurate financial time series forecasting is essential for informed investment and risk management decisions. Traditional methods, including statistical techniques such as SMA, ARIMA, VAR, and LASSO, and deep learning models like LSTM and Transformer, often fall short of capturing the complex seasonal and cyclical dynamics inherent in financial data. To overcome the noted shortcomings, this study introduces a transfer learning approach utilizing an enhanced Transformer model with correlation-based attention mechanisms. This proposed model significantly improves its capacity to capture long-term dependencies and perform robust cross-market predictions. Initially trained on the Dow Jones Index, it demonstrates superior transferability to diverse asset classes, including stock indices, commodities, and cryptocurrencies. Experimental evaluations across multiple metrics, including MSE, MAE, MHD, and R<sup>2</sup>, reveal that the proposed model consistently outperforms benchmarks. Notably, it achieves outstanding predictive accuracy in BTC (MSE: 0.0352) and SET (MSE: 0.1701), establishing a strong foundation for advanced transfer learning applications in financial forecasting across varied markets.

KMITL

Browse

Filters

Settings

Sort By

Results per page

Search Results