A robust deep learning algorithm for lung cancer detection from computed tomography images

Abstract

Detecting lung cancer at its earliest stage offers the best possibility for a cure. Chest computed tomography (CT) scans are a valuable tool for early diagnosis. However, the initial stages of lung cancer may present patterns in the images that are not easily detectable by radiologist, potentially leading to misdiagnosis. Although automated approaches using deep learning (DL) algorithms have been proposed, it depends on a substantial amount of data to achieve diagnostic accuracy comparable to that of radiologists. To alleviate this challenge, this study proposes a DL algorithm that uses an ensemble of convolutional neural networks and trained on relatively small dataset (IQ_OTH/NCCD dataset) to automate lung cancer diagnosis from patient chest CT scans. The method achieved an accuracy of 98.17 %, a sensitivity of 98.21 %, and a specificity of 98.13 % when categorizing scans as either cancerous or non-cancerous. Similarly, it achieved an accuracy of 95.43 %, a sensitivity of 93.40 %, and a specificity of 97.09 % when classifying scans as normal or containing benign or malignant pulmonary nodules. These results demonstrate superior performance compared to previously proposed models, highlighting the effectiveness of DL algorithms for early lung cancer diagnosis and providing a valuable tool to assist radiologists in their assesments.