Development of an AI-based Diagnostic Support System for Radiography Image Analysis

Thesis Abstract

The accuracy and efficiency of radiographic diagnoses are critical to effective patient management, yet the reliance on manual interpretation by radiologists presents significant challenges including fatigue-related errors, variability in expertise, and limited access to specialized knowledge in low-resource settings. This study addresses these issues by developing an AI-based diagnostic support system designed to enhance the precision and speed of radiography image analysis. The primary aim is to create a robust, clinically applicable artificial intelligence framework that facilitates accurate detection and classification of common thoracic and musculoskeletal pathologies. To achieve this, the study sets out specific objectives (1) to curate a comprehensive annotated dataset of radiographic images, (2) to develop and train convolutional neural network (CNN) models for automated image analysis, (3) to evaluate the system’s diagnostic performance against expert radiologists, and (4) to examine the interpretability and integration potential of the AI system within existing radiology workflows. Employing a mixed-methods research design, the quantitative component involves training CNN models using a dataset of 10,000 anonymized radiographic images sourced from a tertiary hospital’s radiology department. These images encompass common thoracic and musculoskeletal conditions, with labels verified by an expert panel. The data collection instruments include digital image repositories, annotation tools, and performance evaluation metrics such as accuracy, sensitivity, specificity, and area under the receiver operating characteristic (ROC) curve. Validation involves k-fold cross-validation and external testing on an independent dataset of 2,000 images to assess generalizability. The qualitative component gathers insights through semi-structured interviews with practicing radiologists and radiology technicians to evaluate system usability and interpretability, analyzed via thematic analysis. Data analysis employs advanced deep learning techniques, primarily convolutional neural networks, optimized through transfer learning with pre-trained models such as ResNet and DenseNet. Performance metrics are statistically analyzed using ROC analysis, paired t-tests for accuracy comparison, and Cohen’s kappa for inter-rater reliability. Furthermore, explainability techniques such as Gradient-weighted Class Activation Mapping (Grad-CAM) are utilized to interpret model decisions and enhance clinical trust. The study also applies the Theory of Technological Acceptance Model (TAM) to explore factors influencing technology adoption among radiology professionals. The anticipated findings suggest that the AI support system achieves diagnostic accuracy comparable to expert radiologists, with sensitivity and specificity exceeding 90% across targeted conditions. The system is expected to significantly reduce the time required for image interpretation, thus improving workflow efficiency. Qualitative insights are likely to reveal high perceived usefulness and ease of integration, although potential barriers such as resistance to change and data privacy concerns may be identified. These results are expected to demonstrate the feasibility and clinical value of integrating AI-driven diagnostic tools into routine radiology practice. This research contributes to knowledge by providing a comprehensive framework for developing, validating, and deploying AI-based diagnostic support systems in radiography, emphasizing both technical performance and contextual acceptability. It advances understanding of how deep learning models can be tailored to specific clinical tasks and offers actionable insights into implementation strategies. The study concludes that well-designed AI systems can augment radiologists’ capabilities, leading to improved diagnostic accuracy and operational efficiency. Recommendations include adopting standardized data protocols, establishing continuous AI model training with diverse datasets, and fostering multidisciplinary collaboration to ensure ethical and user-centered deployment. Future research directions suggested involve extending the system to other imaging modalities, exploring real-time diagnostics, and evaluating long-term clinical outcomes of AI integration in radiology.