TRANSFER LEARNING-BASED LUNG CT IMAGE CLASSIFICATION USING VGG16, RESNET-50, AND NAIVE BAYES: A CORRECTED COMPARATIVE STUDY

Abstract

Lung cancer is the major cause of cancer-related death in the world; early diagnosis is closely correlated with the clinical outcome. While computed tomography (CT) is the most widely used imaging modality for pulmonary nodule assessment, the small size of nodules, inter-reader variability, class imbalance, and the overlap between the radiologic patterns of benign and malignant nodules make visual interpretation challenging. This manuscript involves a corrected and journal-ready comparison between three different models namely Naive Bayes, VGG16, ResNet-50 for the classification of Lung CT images. To eliminate the disease-label inconsistency in the original study, the Study distinguishes data pertaining to pneumonia from classification data for lung cancer, and recommends cancer-specific datasets, such as LIDC-IDRI and IQ-OTH/NCCD. The proposed workflow involves de-identifying the data, splitting the data at the patient level, normalizing the intensity of the CT scans, segmenting the lungs, augmenting the data class-balanced, transferring the data, and evaluating it with accuracy, precision, recall, specificity, F1-score, ROC-AUC, and confidence intervals. The sample size of 20 held out images from a small pilot confusion matrix, with a Naive Bayes model, is only a “sanity check” since this is too small for clinical claims, but it is used. VGG16's accuracy on a six-image pilot set was also reported as 100%, and it was also deemed as non-generalizable. Therefore, the manuscript highlights a repeatable expansion strategy based on public CT datasets and external hospital validation. Authors will be able to submit the corrected protocol for a preliminary comparative study once they have inserted full experimental runs, ethics approval information when required, and source-code availability details.