AN ENHANCED SINGLE-SHOT NEURAL NETWORK FOR DIAGNOSING CHEST DISEASES
Lung diseases pose significant health and social challenges worldwide. X-ray imaging plays a crucial role in diagnosing lung diseases; however, manual analysis is often time-consuming, subjective, and prone to errors. These limitations make manual analysis unsuitable, as timely and accurate diagnoses are essential for effective treatment and management. In this context, Deep Learning (DL) techniques offer significant promise for diagnosing chest diseases. Enhancing the capabilities of the Single Shot Detector (SSD) network and leveraging ensemble methods can revolutionize lung disease diagnosis by addressing gaps in multiclass classification and enabling more precise treatment decisions.
This dissertation focuses on five main objectives:
- Utilizing ensemble methods and lung segmentation to improve the performance of Convolutional Neural Networks (CNNs).
- Effectively employing SSD for detecting six chest diseases.
- Implementing strategies to overcome existing challenges in ensemble methods and enhance ensemble model predictions.
- Exploring the impact of various techniques on the efficiency and accuracy of SSD in object detection.
- Proposing a Refined Object Detector (RefineDet) for lung disease detection by integrating SSD with innovative techniques.
The dissertation introduces four innovative DL models to address these objectives:
- Enhanced Ensemble Model (EEM): Achieves 96.1% accuracy in diagnosing fifteen chest diseases.
- Enhanced SSD Model (ESSDM): Achieves 98.4% accuracy for two specific diseases and 96.5% for six diseases, with prediction times around 0.018 seconds.
- Hybrid CNN and SSD Model (HM): Achieves 98.7% accuracy for two specific diseases and 96.8% for six diseases, with prediction times of approximately 2.3 seconds.
- Enhanced RefineDet Single Shot Detector Model (ERDM): Achieves 99.1% accuracy for two diseases and 97.9% for two and six diverse diseases, with detection times ranging from 0.33 to 1 second.