Lung Cancer Classification using Reinforcement Learning-based Ensemble Learning

Abstract

Lung cancer is a significant health issue affecting millions of people worldwide annually. However, current manual detection methods used by physicians and radiologists to identify lung nodules are inefficient because of the diverse shapes and locations of the nodules in the lungs. New methods are needed to improve the accuracy and speed of detecting lung nodules. This is important because early detection of nodules can increase the likelihood of successful treatment and recovery. This paper introduces a new LLC-QE model that combines ensemble learning and reinforcement learning to classify lung cancer. Initially, the model undergoes pre-training through the utilization of the Artificial Bee Colony (ABC) algorithm. This approach aims to decrease the probability of the model getting stuck in a local optimum. Subsequently, a set of convolutional neural networks (CNNs) is used to simultaneously derive feature vectors from input images, which are subsequently combined for classification in downstream processes. The LIDC-IDRI dataset, predominantly composed of cases without cancer, was employed to train and evaluate the model. To mitigate the dataset imbalance, the training procedure using reinforcement learning is formulated as a series of interconnected decisions. During this process, the images are regarded as states; the network acts as the agent, and the agent is given a greater reward/punishment for accurately/incorrectly classifying the underrepresented class compared to the overrepresented class. The LLC-QE model achieves excellent results (F measure 89.8%; geometric mean 92.7%), outperforming other deep models. Identifying the optimal values for the reward function and determining the ideal number of CNN feature extractors in the ensemble are achieved through experiments conducted on the study dataset. Ablation studies that exclude ABC pre-training and reinforcement learning from the model confirm these components’ independent positive incremental impact on the model’s performance.