Аннотация
Lung cancer represents a significant health challenge both in Kazakhstan and globally, standing out as one of the most fatal forms of cancer. Diagnosis of lung cancer is challenging as symptoms often remain undetectable in the early stages. Furthermore, lung cancer shares clinical features with various other pulmonary conditions, complicating its accurate identification. Accurate diagnosis typically involves lung puncture for subsequent biopsy, a highly invasive and painful procedure for the patient. Therefore, it is crucial to distinguish false positive cases in the diagnostic stage of computed tomography scans. We conducted a comparative analysis of five machine learning models (Logistic Regression, Decision Tree, Random Forest, SVM, and Naïve Bayes Algorithms) based on radiological features extracted from annotated computed tomography scans. We opted for classical machine learning methods because their decision-making process is easier to control compared to neural networks. We evaluated the models in terms of binary and multi-class classification to determine whether a given nodule is related to calcifications or cancers, as well as its classification according to Lung-RADS, enabling the management of whether further biopsy or only routine monitoring is necessary. We used Precision to evaluate the number of False Positive predictions in the binary classification task. Precision emerged as a pivotal metric in our assessment, offering insights into the number of false positive predictions specifically in the binary classification task. For the multi-class classification aspect, we turned to Quadratic Kappa, a robust measure that accounts for the ordinal nature of the Lung-RADS classes. Our analysis was underpinned by a combination of local Kazakhstani data and the publicly available LIDC-IDRI dataset, underscoring our commitment to leveraging diverse data sources to bolster diagnostic capabilities.