Objectives
Lung cancer remains a major global health burden with high mortality rates, primarily due to late-stage diagnosis and limitations in accessible, rapid screening techniques. This study proposes a novel deep learning framework leveraging a hybrid Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Random Forest classification (RF) mode (CNN-LSTM-RF) architecture for the automated detection and classification of lung cancer using radiological Computed Tomography (CT) scan imaging.
Materials and methods
For this study, DICOM (.dcm) images were obtained from The Cancer Imaging Archive (TCIA) containing a total of 355 subjects and a total of 251,135 CT scan images which was divided into 4 major classes of Adenocarcinoma, Small Cell Carcinoma, Large Cell Carcinoma, and Squamous Carcinoma. However, the study uses 249 out of 355 sample patient files containing only lung CT datasets, ignoring all other irrelevant or mixed radiology scans. Moreover, the Large Cell Carcinoma class had a very less sample size (5 patients), and thus was omitted from this study, leaving behind a total sample size of 244 patients which was considered for classification. These slices were normalized and resized to a consistent input dimension after which data augmentation (including rotations, flips, and intensity shifts) was applied to balance the dataset across classes. A custom hybrid CNN was integrated with LSTM for temporal encoding across CT slice sequences, followed by RF classification. At a distribution ratio of 70:20:10, the data was separated into training, testing, and validation datasets.
Results
The custom CNN-LSTM-RF model classified three cancer types – Adenocarcinoma, Small Cell Carcinoma, and Squamous Cell Carcinoma – with a validation accuracy of 97 % and an AUC exceeding 0.95 indicating a well-balanced and highly effective model. In addition to lung cancer classification, histological grading was evaluated among smoking patients, revealing that even aggressive G3 tumors do not occur with excessively higher frequency in smokers, suggesting multifactorial etiologies of lung cancer. Moreover, Early, Intermediate, and Advanced Stage T-Stage tumors were also graphically represented for different age groups. This framework demonstrates promising clinical potential for early-stage lung cancer detection, offering a scalable tool to support radiologists and reduce diagnostic delays.
扫码关注我们
求助内容:
应助结果提醒方式:
