Pattern of Lung Injury in CT/HRCT Using Deep Learning Techniques

Detecting and diagnosing thoraco-pulmonary diseases at an early stage is critical for effective medical intervention. The advent of deep learning has revolutionized the field of medical image analysis, offering unprecedented opportunities for detecting abnormalities from CT images. This paper proposes a neural network based architecture called the MultiResUNet to efficiently analyze the CT scan images and thereby perform accurate and thorough analysis to detect thoraco-pulmonary disorders at an early stage. According to the CT of lungs, there can be four different patterns- Ground glass, Reticulation, Nodules and Mosaic. MultiResUNet has been used to correctly identify the patterns and classify them. The model achieved an average F1 score of 0.874, precision of 0.8745, and an accuracy of 0.9363. Model in all iterations exhibited better or comparable performance for all the evaluated performance metrics with lower standard deviations over the test dataset. A possible future enhancement in this area could be the introduction of an ensemble architecture for classification which has the capability to process 3D images.