Vision-Language Models (VLMs) have advanced rapidly for high-resource languages such as English, yet Bengali remains severely underserved in multimodal AI research. This paper presents a culturally grounded hybrid Bengali image captioning system built upon BanglaVision45K, a dataset of 45,000 images paired with 225,000 human-verified Bengali captions
(five per image), assembled from four complementary sources:
Flickr30k (translated via Google Translate), BNature, Bornon,
and a custom collection of 4,000 manually annotated Bangladeshi
cultural images. One-third of the dataset (33.33%, approximately
15,000 images) represents native Bangladeshi cultural contexts.
The corpus contains 46,499 unique Bengali tokens with an
average caption length of 9.53 words. We propose a novel hybrid
architecture combining a frozen SigLIP2 Vision Transformer
encoder (layers 1–8 frozen, layers 9–12 trainable, patch size 32)
with a custom hybrid decoder integrating 4-layer LSTM, 4-layer
GRU, and BanglaGPT, connected through a Luong multiplicative
attention-based fusion mechanism. Through systematic bench-
marking of eight architectures, SigLIP2 + Att-GRU Ex achieves
the best ablation efficiency-performance trade-off (BLEU-1:
0.5360, METEOR: 0.3821, ROUGE-L: 0.4631, CIDEr: 0.2205).
Our proposed final hybrid model (SigLIP2 + LSTM + GRU +
BanglaGPT) trained on the full BanglaVision45K (90/5/5 split)
achieves state-of-the-art performance: BLEU-1: 0.8356, BLEU-
2: 0.7219, BLEU-3: 0.6119, BLEU-4: 0.5194, METEOR: 0.7004,
ROUGE-L: 0.7414, CIDEr: 0.5287, outperforming all baselines
by over 15% across all metrics.

Abstract—In computational biology, machine learning is now
essential to the analysis of massive transcriptome datasets. Singlecell RNA sequencing (scRNA-Seq) can be used to study glioblastoma multiforme (GBM), the most prevalent and deadly adult
brain tumour, cell by cell. We developed a repeatable, leakagecontrolled machine learning workflow using the GSE84465
dataset (3,589 cells) to categorise each cell according to its surgical origin (tumour core vs. infiltrating periphery) and determine
the genes that differentiate the two compartments.XGBoost, Logistic Regression, Random Forest, AdaBoost, SVM, LightGBM,
Extra Trees, a Neural Network, and a Stacking Ensemble were
the nine classifiers that were assessed. XGBoost was the most
stable under stratified five-fold cross-validation (0.949 ± 0.007),
with held-out accuracy ranging from 0.92 to 0.95 (ROC-AUC
up to 0.993). Immune and microenvironmental factors (e.g.,
CCL3, MIF, ANXA1/ANXA2, HLA-B, VIM) were enriched in
the top 100 XGBoost-ranked genes. Twenty enriched pathways,
twenty hub genes, and fifteen potential medications were obtained
through downstream pathway/GO enrichment, PPI and hub-gene
analysis, TF/miRNA inference, and drug-protein mining. Overall,
the study demonstrates how interpretable machine learning can
transform GBM scRNA-Seq data into prioritised therapeutic
targets and biologically grounded biomarkers.

Abstract—Uterine carcinosarcoma (UCS) is a rare, biphasic,
and highly aggressive gynecologic malignancy with disproportionately poor prognosis. Although microRNAs (miRNAs) regulate
tumorigenesis across many cancers, their prognostic role in
UCS remains largely undefined, owing to the rarity of the
disease and the scarcity of matched expression-plus-survival
cohorts. We sought to derive and internally validate an miRNAbased prognostic signature for overall survival in UCS using
public data. miRNA and mRNA expression and clinical data
for 55 UCS patients (34 deaths, 62%) were obtained from The
Cancer Genome Atlas (TCGA). As weighted correlation network
analysis identified no survival-associated module, we adopted a
direct approach: univariate Cox regression screened survivalassociated miRNAs, and a signature was built by LASSO Cox
regression with the number of miRNAs optimized by repeated
cross-validation. A four-miRNA signature (hsa-miR-1249, hsamiR-1287, hsa-miR-30d, hsa-miR-330) was identified. High-risk
patients had significantly shorter overall survival than low-risk
patients (log-rank P = 6.67 × 10−5). Time-dependent AUCs
were 0.79, 0.74, and 0.76 at 1, 2, and 3 years; repeated 70/30
cross-validation gave a more conservative mean test AUC of 0.69
(95% range 0.41–0.84). The risk score remained an independent
prognostic factor after adjustment for age and stage (hazard ratio
= 2.72, P = 1.67×10−6). Enrichment of correlated target genes
highlighted hypoxia and peptide-hormone responses. This compact signature provides moderate, independent prognostic value
for UCS and may serve as a hypothesis-generating biomarker;
given the small single-cohort sample, external and experimental
validation are required.

The complex assortment of disorders in the human body known as cancer presents numerous difficulties for modern-day medicine. Its molecular mechanisms are so well understood for research advancements in many sectors. Immunotherapies and targeted therapeutics are quite potential for individualized treatments. To enhance early-stage detection and more effective treatments, research in healthcare must continue. In order to evaluate medical imaging data and precisely identify minor patterns indicative of nine types of skin cancer, this research proposes a conceptual framework for diagnosing many skin cancer lesions by utilizing the capabilities of CNN (i.e., Convolutional Neural Networks).Through the usage of the Adam optimizer, the EfficientNetB5 network designs and integrates various adaptive learning rates, allowing it to dynamically adapt and maximizes performance during training. The proposed model’s multi-layered architecture allows it to recognize minute characteristics at microscopic and macroscopic levels, ensuring a comprehension of possible existing cancers. For three hidden layers and output channels of 32, 54, and 128 respectively, the proposed model has the potential to greatly increase diagnosis accuracy and enable rapid treatments. With accuracy of 92.68% for 9 different forms of skin cancer in the dataset, the subsequent modelling assurances class balancing and augmentation using XGBoost classifiers, as it will train the images with better feature handling.

This systematic review synthesizes evidence from 29 peer-reviewed studies published between 2008 and 2025 to evaluate the real-world effectiveness of assistive technologies for the Deaf and Hard-of-Hearing (DHH) community. Although recent advances in artificial intelligence, computer vision, and wearable sensing have produced recognition accuracies exceeding 90% under controlled laboratory conditions, substantial challenges remain in practical deployment. Using a tripartite effectiveness framework, the review found that only one technology was Currently Effective for real-world use, whereas 25 studies were classified as Effective with Further Development because of their dependence on controlled environments. The analysis also identified a significant communication imbalance: most technologies support one-way translation, while only three studies proposed truly bidirectional systems. Furthermore, limited participation of native Deaf signers in the design process often reduced usability and cultural acceptance. These findings highlight the need for user-centered, participatory design, long-term real-world evaluation, and bidirectional communication technologies to achieve meaningful accessibility and social inclusion.

Key Contribution:
● Proposed a two-stage deep learning
approach for automatic detection and
segmentation of bone fractures based on
X-ray images.
● Combined and processed two Kaggle
X-ray datasets in order to get a balanced
dataset for bone fractures detection [7].
● Created a personalized segmented
dataset by doing manual annotation and
masks for fracture localization.
● Preprocessed and augmented the dataset
to increase the quality and
generalization of the deep learning
model [8].
● Used several deep learning models
(ResNet50, EfficientNetB3,
MobileNetV3, Xception, InceptionV3,
and DenseNet121) for detecting bone
fractures.
Designed a
Multi_Bone_Hybrid_DetModel based
on several most accurate detection
models to get better detection results.
● Used several U-Net-like segmentation
models and created a
Multi_Bone_Hybrid_SegModel to get
better fracture localization results.
● Added Grad-CAM for both detection
and segmentation models to improve
explainability and interpretation of the
model [9].
● Created a web-based application for
fracture detection, segmentation, and
visualization of Grad-CAM

The contribution of research in lip reading using deep
learning model named modified dual-input 3D Convolutional
Neural Network (3DCNN) is fundamentally transforming
communication accessibility for individuals with hearing
impairments. Additionally, we explore some models like as
3D Convolutional Neural Network (3D-CNN), 2D
Convolutional Neural Network with GRU, 2D Convolution
Neural Network with LSTM (2DCNN with LSTM) on our
dataset and evaluate the performance and finally compare the
performances with our proposed system

Deep learning-based Monkeypox (Mpox) skin lesion
screening offers a rapid and low-cost decision-support solution
for resource-constrained healthcare settings. However, existing
approaches often depend on limited datasets, computationally
intensive architectures, and augmentation procedures that may
introduce data leakage. This paper proposes EE-MpoxLiteNet,
a lightweight, edge-enhanced, and explainable framework for
Mpox skin lesion classification. A leakage-safe protocol first
partitions the original images into training, validation, and
test subsets before augmentation. The proposed Edge-Aware
Lesion Enhancement Module combines contrast-limited adap-
tive histogram equalization, bilateral filtering, and Sobel-based
structural information to improve lesion contrast and boundary
representation. A dual-stream architecture then captures com-
plementary appearance-based semantic and edge-guided mor-
phological features. These representations are refined through
a Multi-Scale Lesion Representation Block and a Cross-Branch
Attention Fusion Module before classification. Experiments on
the Monkeypox Skin Lesion Dataset show that EE-MpoxLiteNet
achieves 97.8% accuracy, a 97.8% F1-score, and a 99.4% ROC-
AUC, outperforming standard CNN baselines and heavier hybrid
models. The network contains 6.85 million parameters and
requires only 0.86 GFLOPs, demonstrating a favorable accuracy–
efficiency trade-off. Grad-CAM visualizations further show that
predictions are primarily based on lesion-relevant regions. These
findings support the potential of EE-MpoxLiteNet for efficient
and explainable AI-assisted Mpox screening.

The main contributions of this work are summarized as
follows:
1) We create a kind of framework that combines three
things: EfficientNetB1, ResNet50 and MobileNetV2.
This combination helps the framework to work when
it comes to classifying diseases in rice leaves that look
very similar.
2) We try out six pre-trained networks. VGG19, ResNet50,
DenseNet121, MobileNetV2, EfficientNetB0 and EfficientNetB1. To see which one works best. We keep
everything the same like how we prepare the data and
how we train the networks.
3) We also test the trained models using pictures of rice
leaves that were taken in the field. These pictures were
taken in light with different backgrounds and, with
different cameras. This helps us see how well the models
can work in life when the conditions are not perfect.
4) To make the models more understandable we use something called Grad-CAM visualization. This helps us see
which parts of the leaf the model is looking at when it
makes a prediction.

Dental caries is a widespread oral disease whose
early detection remains difficult because conventional visual
assessment is subjective, time-consuming, and prone to examiner variability. This study proposes an automated detection
framework that combines deep residual feature extraction with
bidirectional multi-scale feature aggregation to improve localization of small and low-contrast carious lesions in intraoral
photographs. Transfer learning, image normalization, and geometric and photometric augmentation are employed to enhance
robustness. The framework is evaluated on the publicly available
Annotated Intraoral Image Dataset for Dental Caries Detection
and compared with several alternative backbone architectures.
Experimental results show a precision of 0.8173, recall of 0.7601,
F1-score of 0.7877, mAP@0.5 of 0.7998, and mAP@0.5:0.95
of 0.6361. Image-level evaluation further yields a sensitivity
of 0.8894, specificity of 0.9502, and accuracy of 0.9283. These
findings demonstrate effective lesion localization and clinically
relevant screening performance while providing a reproducible
benchmark for future research on reliable computer-aided dental
caries detection.

This paper proposes CIDR, a three-stage framework integrating entropy-based drift detection, influence-guided targeted unlearning, and conformal recalibration to restore model reliability under distribution shift. CIDR restores coverage to 0.903 ± 0.034 against a target of 0.90 with no statistically significant accuracy degradation, outperforms full retraining in accuracy (0.770 vs. 0.669) and efficiency (30% smaller prediction sets, 1/10 the cost), and provides a practical solution for maintaining trustworthy AI under drift.

Software testing remains one of the most labor-intensive stages of the software development lifecycle, and conventional automation frameworks rely on brittle, manually written scripts that require constant upkeep as applications evolve. The recent emergence of large language models (LLMs) capable of understanding natural-language requirements and producing syntactically correct code has opened a new path toward intelligent test automation. This paper synthesizes current research and proposes a layered architecture, referred to as the LLM-Driven Test Automation Framework (LTAF), that integrates retrieval-augmented context ingestion, LLM-based test-case generation, automated script synthesis for common automation stacks, a self-healing locator component, and a closed feedback loop between test execution and the generative model. We describe the architecture and design rationale of each component, discuss how the framework could be evaluated against established baselines such as search-based test generation, and summarize the reported strengths and limitations of LLM-based testing approaches drawn from the current body of work. We further identify open challenges, including the test-oracle problem, hallucinated assertions, flaky-test diagnosis, and data-privacy concerns when proprietary code is shared with external model providers, and we outline directions for future research toward reliable, self-repairing test automation pipelines.