Chronic Lung Disease Identification in Chest CT Scans Through Hybrid Deep Transfer Learning Framework

Abstract

Accurate and timely diagnosis is essential for successful treatment of chronic pulmonary conditions, particularly lung tumours. To capture more complicated pathological characteristics, diagnostic algorithms based on current data sets are restricted to using older databases and single branch systems. This research introduces a new solution: A hybrid deep learning model applied to a recently released quantitative database of CT scans of the chest. This method combines a pre-trained ResNet50 backbone with a Multi-Layer Perceptron (MLP) to merge high level spatial properties with statistically determined measures (mean intensity, standard deviation and entropy). In addition, the two branch fusion handles extremely mechanical diagnosis and global spatial characteristics of the images, facilitating the classification of adenocarcinoma, large cell carcinoma, squamous cell carcinoma and normal tissue. After extensive preprocessing and data augmentation, our framework reached a reliable training accuracy of 94.00% and a testing accuracy of 88.0%. While this represents the greater benefit of incorporating both deep spatial learning into statistical measures, this specific architecture was purposely optimized for functioning within standard computational systems that are restricted to a limited amount of CPU and memory resources. As such, this work serves as a resource-efficient proof-of-concept. This framework provides a foundation for a potential computer-aided diagnostic support system that could assist radiologists in identifying chronic lung conditions.