Azurah A. Samah

Skeletal malocclusion, a common orthodontic condition, affects jaw function and dental health. It is often caused by genetic factors, abnormal growth, bad habits, or trauma. Conventional diagnostic models often fail to generalize across diverse datasets, leading to overfitting and poor test performance. This study aimed to improve diagnostic accuracy by incorporating Vision Attention mechanisms into a custom Convolutional Neural Network (CNN), enabling the model to focus on critical regions in X-ray images. A total of 491 radiographic images depicting facial skeletal structures with various malocclusion types (Classes 1, 2, and 3) were used in this study. A custom CNN was developed and evaluated both with and without attention mechanisms—specifically, Scaled Dot Product Attention and Multihead Attention—to assess their impact on classification performance. The baseline CNN without attention achieved an accuracy of 0.68. With Scaled Dot Product Attention, accuracy improved to 0.72, while Multihead Attention achieved the highest accuracy of 0.76. Evaluation using weighted average precision, recall, and F1-score showed that attention mechanisms significantly enhanced the model’s ability to differentiate between malocclusion classes. Notably, the Multihead Attention model yielded the best performance, reducing misclassification errors and improving generalization. Confusion matrix analysis revealed that it had the lowest classification errors, especially in distinguishing between Class 0 and Class 1. These results suggest that incorporating attention mechanisms, particularly Multihead Attention, enhances CNN performance by improving feature extraction and classification accuracy. Future research should explore more diverse datasets and implement advanced augmentation techniques to improve clinical reliability.

Medical imaging has been essential and has provided clinicians with useful information about the human body to diagnose various health issues. Early diagnosis of diseases based on medical imaging can mitigate the risk of severe consequences and enhance long-term health outcomes. Nevertheless, the task of diagnosing diseases based on medical imaging can be challenging due to the exclusive ability of clinicians to interpret the outcomes of medical imaging, which is time-consuming and susceptible to human fallibility. The ensemble model has the potential to enhance the accuracy of diagnoses of diseases based on medical imaging by analyzing vast volumes of data and identifying trends that may not be immediately apparent to doctors. However, it takes a lot of memory and processing resources to train and maintain several ensemble models. These challenges highlight the necessity of effective and scalable ensemble models that can manage the intricacies of medical imaging assignments. This study employed an SLR technique to explore the latest advancements and approaches. By conducting a thorough and systematic search of Scopus and Web of Science databases in accordance with the principles outlined in the PRISMA, employing keywords namely ensemble model and medical imaging. This study included a total of 75 papers that were published between 2019 and 2024. The categorization, methodologies, and use of medical imaging were key factors examined in the analysis of the 30 cited papers included in this study, with a focus on diagnosing diseases. Researchers have observed the emergence of an ensemble model for disease diagnosis using medical imaging since it has demonstrated improved accuracy and may guide future studies by highlighting the limitations of the ensemble model.

Genetic and genomic variations are primary drivers of tumor development. Identifying driver genes from numerous passenger genes across pan-cancer poses a significant challenge due to varying mutation loads. While independent studies have elucidated cancer-associated mutation patterns within specific cancer types, a systematic approach to integrating these mutation data for assessing the impact of gene mutations has been lacking. This study addresses this gap by integrating pan-cancer genomic somatic mutation data and introducing a novel mutation weight fusion (WeiFu) score calculation method. WeiFu computes frequency and weighted fusion scores by cancer type, facilitating the identification of potential driver genes. Evaluation results on an integrated pan-cancer dataset comprising 29 different cancer types demonstrate that WeiFu significantly outperforms current well-known approaches in prediction accuracy, sensitivity, and specificity. Notably, WeiFu recovers 277 known cancer genes among the top 500 ranked candidates and successfully identifies potential driver genes supported by strong evidence. Consequently, WeiFu shows considerable promise for identifying driver genes within the rapidly expanding corpus of cancer genomic data.

Biological data obtained from sequencing technologies is growing exponentially. Multi-omics data is one of the biological data that exhibits high dimensionality, or more commonly known as the curse of dimensionality. The curse of dimensionality occurs when the dataset contains many features or attributes but with significantly fewer samples or observations. The study focuses on mitigating the curse of dimensionality by implementing Support Vector Machine – Recursive Feature Elimination (SVM-RFE) as the selected feature selection method in the lung cancer (LUSC) multi-omics dataset integrated from three single omics dataset comprising genomics, transcriptomics and epigenomics, and assess the quality of the selected feature subsets using SDAE and VAE deep learning classifiers. In this study, the LUSC datasets first undergo data pre-processing, including checking for missing values, normalization, and removing zero variance features. The cleaned LUSC datasets are then integrated to form a multi-omics dataset. Feature selection was performed on the LUSC multi-omics data using SVM-RFE to select several optimal feature subsets. The five smallest feature subsets (FS) are used in classification using SDAE and VAE neural networks to assess the quality of the feature subsets. The results show that all 5 VAE models can obtain an accuracy and AUC score of 1.000, while only 2 out of 5 SDAE models (FS 1000 & 4000) can do so. 3 out of 5 SDAE models have an AUC score of 0.500, indicating zero capability in separating the binary class labels. The study concludes that a fine-tuned supervised learning VAE model has better capability in classification tasks compared to SDAE models for this specific study. Additionally, 1000 and 4000 are the two most optimal feature subsets selected by the SVM-RFE algorithm. The SDAE and VAE models built with these feature subsets achieve the best classification results.