روژين ستايشي

This thesis examines the application of artificial intelligence, particularly deep learning, in the diagnosis of lung diseases such as tuberculosis and lung cancer in bronchoscopy images. Lung diseases, including tuberculosis and lung cancer, are among the leading causes of death worldwide and present significant challenges in diagnosis and treatment. Two distinct system are presented in this thesis. The first is a multi-task model based on DenseNet-121 with two classifiers. It simultaneously performs image classification as normal or abnormal using a squeeze and excitation module and, in the case of an abnormal detection, classifies the type of anomaly (such as tuberculosis or cancer, which often share overlapping symptoms). This system, eva‎luated on a dataset of 515 images, achieved an overall accuracy of 90.6%, with sensitivity rates of 91.3% for cancer, 81.5% for tuberculosis, and 96.2% for normal cases. The second system proposes a cascading system based on DenseNet-121, consisting of two separate models: the first is a normal-abnormal classifier to detect abnormal findings in bronchoscopy images, and the second is a tuberculosis-cancer classifier to differentiate between tuberculosis and cancer cases. To improve performance and reduce computational costs, model distillation techniques were used to transfer knowledge from complex models to smaller, optimized versions. In the distilled cascading system, the number of dense layers in the normal-abnormal classifier was reduced from 58 to 34, and in the tuberculosis-cancer classifier from 58 to 32. Despite this reduction, the system was able to improve its diagnostic accuracy. This cascading system, eva‎luated on a similar dataset, achieved an overall accuracy of 91.5%, with sensitivity rates of 92.2% for cancer, 80.3% for tuberculosis, and 98.8% for normal cases. This thesis demonstrates the effectiveness of AI-based diagnostic systems in addressing the limitations of current diagnostic methods, optimizing medical image analysis, improving diagnostic accuracy, increasing detection speed, reducing computational costs, and providing more effective support to physicians in diagnosing lung diseases during bronchoscopy. These result in improved healthcare services, reduced human errors, and fewer unnecessary invasive procedures during bronchoscopy.

سيستم كمك تشخيصي رايانەاي , ، سيستم چندوظيفەاي , ، سيستم آبشاري , ، تقطير مدل , شبكه DenseNet , ماژول فشردەسازي و تحريك , برونكوسكوپي , سرطان ريه , سل

CADx , Multi-task system , Cascaded system , Model distillation , DenseNet models , Squeeze and excitation module (SE) , Bronchoscopy , Lung cancer , Tuberculosis

Rojin Setayeshi

Javad Vahidi