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

Thesis Abstract

The accurate and timely diagnosis of medical conditions through radiographic image analysis remains a critical challenge in clinical radiology, often constrained by the limitations of human interpretation variability and increasing workload pressures on radiologists. This study aims to develop an artificial intelligence (AI)-based diagnostic support system designed to enhance the accuracy, efficiency, and consistency of radiographic image interpretation. The primary objectives are to investigate existing AI techniques applicable to radiographic analysis, design a bespoke machine learning framework leveraging convolutional neural networks (CNNs), and evaluate its diagnostic performance against traditional methods. The research adopts a mixed-methods approach, combining quantitative model development and evaluation with qualitative assessments of radiologist acceptance and usability. The study is conducted within a tertiary hospital radiology department, with a target population comprising 2000 anonymized radiographic images of chest X-rays, pelvic radiographs, and extremity images collected from hospital archives. A stratified random sampling technique selected a subset of 500 images representing common pathologies such as fractures, lesions, infections, and tumors. Data collection involves the compilation of annotated image datasets verified by expert radiologists, alongside structured questionnaires and interview guides assessing radiologist perspectives on AI integration. Data preparation includes preprocessing steps like normalization, augmentation, and feature extraction. The core of the methodology involves designing a CNN architecture based on transfer learning principles, utilizing pre-trained models such as ResNet50 and InceptionV3, fine-tuned on the domain-specific radiographic dataset. The model’s diagnostic accuracy is validated through 10-fold cross-validation, and performance metrics including sensitivity, specificity, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC-ROC) are computed. Comparative analyses are conducted with existing AI tools and expert radiologist diagnoses employing paired t-tests and Bland-Altman analysis to assess agreement levels. The study also applies the Technology Acceptance Model (TAM) and the Unified Theory of Acceptance and Use of Technology (UTAUT) frameworks to analyze qualitative data on user perceptions. Expected findings indicate that the AI system will attain a diagnostic accuracy exceeding 90%, with significant improvements in speed and consistency compared with conventional human interpretation. The AI’s ability to detect specific pathologies is anticipated to demonstrate high sensitivity and specificity across varied radiographic modalities. Furthermore, qualitative insights are expected to reveal high acceptance levels among radiologists, citing enhanced diagnostic confidence and workflow efficiency as primary advantages. This research is projected to contribute novel insights into the integration of deep learning models within clinical radiology, filling existing gaps concerning scalability, domain adaptability, and user acceptance. The innovative application of transfer learning tailored to radiography is expected to advance current AI diagnostic tools, offering a practical and robust solution for resource-limited settings. The main conclusion underscores that AI-based diagnostic support systems can substantially augment radiologists’ capabilities, enabling quicker and more accurate diagnoses that potentially improve patient outcomes. Recommendations include adopting the developed AI system within broader clinical workflows, implementing targeted training programs for radiologists, and establishing continuous learning cycles for model updates. Future research should focus on longitudinal assessments of AI system impacts on clinical decision-making, extending to other imaging modalities such as MRI and CT scans, and exploring integrated multi-modal diagnostic models. This study advocates for the systematic adoption of AI-driven solutions as integral components of modern radiology departments to enhance diagnostic quality and operational efficiency.