DSpace Community:

Describing Land Cover Changes via Multi-Temporal Remote Sensing Image Captioning Using LLM, ViT, and LoRA

2026-01-04T00:00:00Z

Title: Describing Land Cover Changes via Multi-Temporal Remote Sensing Image Captioning Using LLM, ViT, and LoRA Authors: Lamar-Leon, Javier; Nogueira, Vitor; Salgueiro, Pedro; Quaresma, Paulo Editors: Pan, Jiayi; Li, Xinghua Abstract: Describing land cover changes from multi-temporal remote sensing imagery requires capturing both visual transformations and their semantic meaning in natural language. Existing methods often struggle to balance visual accuracy with descriptive coherence. We propose MVLT-LoRA-CC (Multi-modal Vision Language Transformer with Low-Rank Adaptation for Change Captioning), a framework that integrates a Vision Transformer (ViT), a Large Language Model (LLM), and Low-Rank Adaptation (LoRA) for efficient multi-modal learning. The model processes paired temporal images through patch embeddings and transformer blocks, aligning visual and textual representations via a multi-modal adapter. To improve efficiency and avoid unnecessary parameter growth, LoRA modules are selectively inserted only into the attention projection layers and cross-modal adapter blocks rather than being uniformly applied to all linear layers. This targeted design preserves general linguistic knowledge while enabling effective adaptation to remote sensing change description. To assess performance, we introduce the Complementary Consistency Score (CCS) framework, which evaluates both descriptive fidelity for change instances and classification accuracy for no change cases. Experiments on the LEVIR-CC test set demonstrate that MVLT-LoRA-CC generates semantically accurate captions, surpassing prior methods in both descriptive richness and temporal change recognition. The approach establishes a scalable solution for multi-modal land cover change description in remote sensing applications.

Are Small Language Models Enough for Biomedical QA Tasks?

2025-08-17T23:00:00Z

Title: Are Small Language Models Enough for Biomedical QA Tasks? Authors: Lamar-Leon, Javier; Nogueira, Vitor; Quaresma, Paulo Abstract: This paper presents a specialized fine-tuning approach for the Mistral-7B Large Language Model (LLM) tailored for biomedical applications. We employ Low-Rank Adaptation (LoRA), a parameter-efficient fine-tuning method, to adapt the model to the intricacies of biomedical language and domain-specific knowledge. By integrating LoRA, we aim to preserve the general language understanding capabilities of Mistral-7B while enhancing its performance on biomedical tasks. The fine-tuning process involves training the model on the PubMedQA dataset. Our experiments demonstrate that the fine-tuned Mistral-7B model achieves notable accuracy, 60%. This performance is particularly significant given the relatively modest size of the Mistral-7B model compared to other approaches that often require larger models to achieve comparable results. The results highlight the effectiveness of LoRA in fine-tuning large language models for domain-specific applications, particularly in the biomedical field, where precise and contextually accurate language understanding is crucial. This work contributes to the advancement of AI in healthcare by providing a robust and efficient method for adapting LLMs to biomedical applications, demonstrating that high precision can be achieved with a smaller model size.

Feature Extraction of Apparent Diffusion Coefficient in Human Brain Lesions to Distinguish Benign and Malignant Using MRI

2023-03-01T00:00:00Z

Title: Feature Extraction of Apparent Diffusion Coefficient in Human Brain Lesions to Distinguish Benign and Malignant Using MRI Authors: Vijithananda, H.H.T.S.M. Abstract: Introduction: Apparent Diffusion Coefficient (ADC) is one of the most common Magnetic Resonance Imaging (MRI) techniques that is frequently used in the brain tumor identification process in the current clinical neuro-imaging setup. The ADC quantitatively measures the diffusivity of water molecules within living tissues using Diffusion-Weighted Imaging (DWI) and provides information about the net direction of the water diffusion and the boundaries that restrict the diffusion which are crucial to identify certain pathological conditions in the cellular level. Objectives: This study focused on extracting image texture features from MRI-ADC images of human brain tumors, and the patient demographics, identifying the distribution pattern of each feature, and developing robust Machine Learning (ML) models that categorize the tumors according to benign or malignant nature as well as the aggressiveness of gliomas. Materials and Methods: This prospective study was designed to be conducted in three main steps: identify the key features that correlate with tumor types through a basic statistical analysis, develop a ML model that predicts the benign and malignant nature of the tumor, and developing an ML model to classify gliomas according to the aggressiveness of tumors. The study was carried out using 1790 MRI-ADC brain image slices (980 malignant, 8 benign) from 252 human subjects including both males and females (age 2 years to 90 years) who were radiologically and histopathologically diagnosed with brain carcinoma. All MR images were acquired utilizing 3T MR systems using a head coil. ADC images of brain tumors were generated by merging b=0 and b=1000 diffusion- weighted images. Pixels within the tumor region of the generated ADC images were selected by drawing a region of interest (ROI) surrounding the tumor area. The features; i.e., mean pixel value, higher-order moments (skewness, kurtosis) of ADC, Grey Level Co-occurrence Matrix (GLCM) texture features, patient’s age, and gender corresponding to each ROI were extracted. The extracted features were tested with a one-tailed P-value hypothesis testing with a 95% confidence level by hypothesizing that there is no significant difference in mean values of extracted features among benign and malignant brain tumors. Furthermore, two ML classification models were developed using the extracted features: a classification model to differentiate benign and malignant brain tumors and a classification model to differentiate gliomas within the dataset according to the World Health Organization (WHO) glioma grading system. Development of ML model to differentiate benign and malignant image slices: The data extracted from 1549 image slices of 205 subjects with brain tumors (excluding pituitary macroadenoma and dermoid cysts from the initial dataset) was split into training (70%) and test (30%) sets. The analysis of variance (ANOVA) f-test feature selection over the train set was utilized to select the best set of features to train an ML model, and the K(10)-fold cross-validation method was utilized to find the most promising ML algorithm and corresponding hyper-parameters over the training dataset. The hyper-parameters of the algorithm were tuned using a grid search technique and the decision threshold was adjusted to have the optimum level of classification power. The performance of the tuned model for benign and malignant tumor classification was assessed using the accuracy measure over the test set. Development of ML classification model to differentiate WHO glioma grade: The glioma classification model was developed based on 1088 labeled MRI-ADC glioma brain image slices acquired from 88 human subjects. The gliomas were categorized into four groups according to the WHO glioma grading system (WHO-I, WHO-II, WHO-III, and WHO-IV). The glioma dataset was split into train and test sets with 70% to 30% proportions respectively and the best set of features to build a predictive ML model was selected by applying the ANOVA f-test over the train set. The most promising supervised learning ML algorithm for the glioma dataset was selected using the K-fold cross-validation and the hyper-parameters of the developed classification model were optimized using the grid search technique. Finally, the performance of the tuned glioma classification model was assessed using the accuracy measure over the test set. Results: According to the P-values obtained from each feature the mean pixel value of ADC and GLCM texture features i.e., mean1, mean2, variance1, variance2, energy, and contrast showed significantly (P-value < 0.05) higher feature values for benign tumors while the kurtosis and GLCM texture features i.e., entropy, homogeneity, correlation, prominence, and shade showing significantly high feature values for malignant tumors. However, facts for the features; skewness (P-value 0.05 < 0.0603), and the patient’s age (P-value 0.05 < 0.2729) were not enough to reject the null hypothesis of this study. According to the ANOVA f-test conducted over the train set of the benign and malignant dataset, the skewness expressed the minimum ANOVA f-test score (0.8731). The Random Forest Classifier (RFC) algorithm was selected to build the benign and malignant brain tumor classification model as it scored the highest mean cross-validation score (0.8899±0.0217) at the K-fold cross-validation experiment. The classification model developed in RFC was able to distinguish benign from malignant brain tumors with an accuracy of 91.16% over the test set (after the hyper-parameter tuning process). According to the ANOVA f-test, two attributes: the GLCM energy (14.21), and correlation (21.78) performed minimum scores and were excluded from the glioma dataset. Among the tested algorithms, the RFC (0.8536 ± 0.0199) obtained the highest score and was selected to build the glioma prediction model with an accuracy of 84.80% over the test set. Conclusion: The above three experiments revealed the feasibility of the utilization of texture features of MRI-ADC images for tumor classification. Therefore, this study’s outcomes enable the development of advanced tumor classification applications that assist in the decision-making process in a real-time clinical environment.

A Blockchain Approach to IoT Sensor Data Storage Using Hyperledger Fabric

2025-08-31T23:00:00Z

Title: A Blockchain Approach to IoT Sensor Data Storage Using Hyperledger Fabric Authors: Solipa, Diogo; Saias, Jóse; Salgueiro, Pedro Abstract: The integration of IoT and blockchain technologies presents a compelling opportunity to address longstanding challenges related to data integrity, trust, and decentralization in distributed systems. blockchain is a specific implementation of distributed ledger technologies (DLTs), known for its security, trust, and decentralization. IoT, on the other hand, has become increasingly relevant, generating vast amounts of data and requiring high uptime for connected devices, which are susceptible to data tampering, privacy breaches, and single points of failure. This paper explores the integration of blockchain with an IoT network. Particular emphasis is placed on Hyperledger Fabric, a permissioned blockchain framework tailored for enterprise-grade applications. Through a combination of architectural analysis and empirical testing, this work evaluates multiple Hyperledger Fabric configurations and performance metrics such as response time, throughput, latency, scalability. The experimental component leverages benchmarking tools, such as Hy- perledger Caliper. These tests aim to quantify trade-offs between query flexibility and system performance under varying configurations and work- loads. The findings support the viability of using tailored blockchain con- figurations for scalable and secure IoT deployments, while highlighting areas for further optimization and future work.