Final Thesis: AI Project in Healthcare: Utilizing Computer Vision for Medical Images

Abstract: The rising rates of respiratory diseases like pneumonia and COVID-19 have made the need for fast, reliable, and accessible tools to analyze medical images important and necessary. Chest X-rays (CXRs) are still one of the most prevalent imaging modalities to be used in the diagnosis of these conditions, but they normally require expert radiological evaluation, which in certain situations may not be conveniently accessible in underprivileged conditions. This thesis addresses the use of deep learning methods in the automatic interpretation of CXR images in aiding the process of lung disease detection over healthy images. Several convolutional neural network (CNN) models were tested against each other, such as VGG, ResNet. ResNet-50 was used as a transfer learning model, and several ensemble mechanisms were tried. ResNet-50 and VGG-16 combined with equal weights as an ensemble, had produced the best results. Interestingly, in all models, training was performed on image-level labels only (excluding annotated lesion coordinates), which shows that high performance can be achieved even without detailed annotations. The accuracy of the final ensemble model is 94.98%, precision is 95.26%, recall is 94.98%, and the F1-score is 94.96%. It has, as well, an AUROC value of 0.9812 and a 97.37% specificity, which emphasize its high discriminative ability. A graphical user interface (GUI) has been designed to implement usability, requiring the user to enter the symptoms and also to upload a CXR image. Both image-based and symptom-based diagnostic predictions are returned by the system, and this aspect means that the system can be fully integrated into clinical workflows. Future directions will involve further enhancement in the interpretability of models, classification of other thoracic disorders, and evaluation of the system in practice in clinical practice.

Keywords: None

PDF: Master Thesis

Reference: Mit Ashokbhai Desai. AI Project in Healthcare: Utilizing Computer Vision for Medical Images. Master Thesis. Friedrich-Alexander-Universität Erlangen-Nürnberg: 2025.