OBJECTIVES : The purpose of our review was to assess the performance of ML algorithms in detecting lung cancer and compare it against healthcare professionals.

METHODS : In this review, we searched for peer-reviewed journal publications that developed an ML model for lung cancer diagnosis and additionally compared the performance indices made by algorithms and physicians, including radiologists. Due to different cutoff thresholds, the meta-analysis using inverse variance weighting was conducted with the ROC curve metric only.

RESULTS : A total of 11 studies were identified that compared the sensitivity, specificity, diagnostic effectiveness, and/or AUC. These studies retrospectively collected data and were published between 2017 and 2020. The method for predictor measurement was consistent with 9/11 (82%) studies investigating computed tomography and one study, each inspecting chest radiograph, and histopathology images. Seven studies used varying models of expert consensus, and four used histopathology as reference standards. Within deep learning models, all studies evaluated the performance of the convolutional neural network. The sensitivity and specificity of neural networks ranged 0.628-0.96 and 0.778-0.98, while clinicians reported 0.683-0.90 and 0.628-0.993, respectively. The average diagnostic effectiveness of human observation was 75.45%, which improved to 81.05% with the aid of neural networks. The AUC scores ranged from 0.771-0.997, while the pooled-effect estimates were 0.925 (95% CI: 0.924-0.926) and 0.907 (95% CI: 0.858-0.956) for fixed and random effects models, respectively. The high AUC value of 0.907 (preferred due to high I²) signifies neural networks an excellent classifier, with a 90.7% chance that it will correctly distinguish a normal from a lung cancer affected patient.

CONCLUSIONS : Based on the findings of our review and exploratory meta-analysis, we can cautiously state that the diagnostic performance of neural networks is at least equivalent to healthcare experts. However, additional studies considering the integration of such algorithms are needed to support the implementation of this promising technology further.