Segmentation of Tuberculosis-Affected Lungs from Chest X-Ray Images using Fine-Tuned Transfer Learning Model

Lung disease is one of the largest contagious diseases which is the main cause of death in children and adults across the world. Lung disease can degrade the breathing ability and pulmonary functionality of the lungs. Viral, bacterial or fungal infections cause these diseases. The objective of this study is to introduce a modified version of the deep learning segmentation model which supports a smaller number of datasets and compares its evaluation parameters with the conventional deep learning model. It shows that the classification of lung X-ray images followed by segmentation improves the model accuracy. In the proposed model, first, we trained U-Net model for the segmentation of lungs on a chest X-ray dataset of Covid disease and saved this model. Then took this model and with the help of a transfer learning approach, we trained this model on a limited tuberculosis dataset. As a result, the proposed model helps to improve accuracy from 68.32% to 94.8% for TB. Other parameters such as dice coefficient, IoU and loss have also been improved in the proposed model.