Title : Artificial intelligence based early detection and subtype classification of histopathologically confirmed lung cancer using thoracic CT images
Abstract:
Objective: Lung Cancer (LCa) ranks first in cancer-related mortality. It is essential to quickly determine the diagnosis and subtype of LCa, which will affect patients' treatment and life expectancy. Currently, diagnosis is made only by histopathological examination. In recent years, studies on artificial intelligence (AI) have been increasing in diagnosing many diseases, pathological and molecular classification of diseases, and prognosis prediction. In this study, we retrospectively examined the diagnosis of LCa and its differentiation into sub-groups using computer-aided systems.
Methods: Approval was obtained from the Malatya Turgut Özal University Ethics Committee for this study (Approval No.: E-30785963-020-170163; dated August 18, 2023). The study was conducted at Malatya Training and Research Hospital. As this was a retrospective study, informed consent was not required. We retrospectively evaluated patients with a pathological diagnosis and histological subtyping of LCa who had a mass detected on thoracic CT and subsequently underwent biopsy at our hospital between January 2014 and December 2022, as well as healthy adults with completely normal thoracic CT findings. The study included patients aged ≥18 years who were pathologically diagnosed with the two most common subtypes of Non-small cell lung cancer (NSCLCa) (adenocarcinoma, squamous cell carcinoma) and Small cell lung cancer (SCLCa) and had a pulmonary mass detected on thoracic CT. We used 864 thorax computed tomography (CT) images. In our study, images of patients diagnosed with the LCa and who had a mass detected on thorax CT and randomly selected people with normal CT were used. This study proposed a new model called LungCTNet to classify thorax CT images. In the second stage of our study, Denoising Convolutional Neural Networks (DnCNN) were used to remove noise in thorax CT images and improve the quality of the images.
Results: The proposed artificial intelligence-based LungCTNet model achieved an accuracy of 98.3% in the detection and histological subtype classification of lung cancer using thoracic CT images. These findings demonstrate the high diagnostic and classification performance of the proposed model and suggest its potential utility as a non-invasive imaging-based approach for lung cancer detection and subtyping.
Conclusion: Cancer cases are increasing day by day in our country and around the world. The most important step for starting treatment is the histopathological diagnosis of the disease. However, since a biopsy cannot be performed on every patient or the biopsy result cannot always be reported quickly and accurately, difficulties in diagnosing LCa continue and treatments cannot be started. According to the findings of our study, LCa can be diagnosed and subtyped with AI, a non-invasive method, without biopsy with the models we have suggested. As a result, a significant portion of patients, especially those who are elderly and whose tissue diagnosis cannot be made because biopsy cannot be performed due to additional diseases such as chronic pulmonary, heart can be diagnosed with AI. Once the diagnosis is made, these patients with poor survival can receive treatment and we can significantly increase their survival due to LCa.
Keywords: Artificial Intelligence, Classification, CT, Deep Learning, Lung Cancer
