ABSTRACT

Medical imaging is crucial for disease diagnosis, clinical evaluation, and treatment planning. Deep learning is increasingly used for medical image segmentation and classification due to the availability of datasets and advancements in medical imaging technology. The study focuses on the impact of classification without segmentation on classification accuracy.

This thesis will create kurd-covid dataset with a ground truth mask from Kurdistan health centers and develop a Deep Learning model to accurately segment lung regions in medical x-ray images. Prove the segmentation is the significant step for medical diagnosis.

This study proposes a deep learning system to accurately segment lung regions in chest x-ray images using a modified U-Net architecture trained on three datasets, including a newly created kurd-covid dataset. The system is then used to improve COVID-19 detection by classifying images to COVID and normal using pre-trained CNN models.

The proposed segmentation model achieved excellent performance on the covid-Qu dataset (97.05% IoU, 99.59% DSC) and the kurd-covid dataset (71.84% IoU, 82.2% DSC). Segmentation improved COVID-19 detection in medical images. InceptionV3 accuracy increased from 96.53% to 96.92% for the kurd-covid dataset and from 93.25% to 98.25% for the covid-Qu dataset. ResNet50 accuracy increased from 95.76% to 98.47% for the kurd-covid dataset and from 91.75% to 98.75% for the covid-Qu dataset.”

In conclude Thesis showed the importance of segmentation in medical image analysis. New KURD-covid dataset and Deep Learning model improved COVID-19 detection via accurate lung region segmentation.