Augmenting lung cancer diagnosis on chest radiographs: positioning artificial intelligence to improve radiologist performance.

AIM: To evaluate the role that artificial intelligence (AI) could play in assisting radiologists as the first reader of chest radiographs (CXRs), to increase the accuracy and efficiency of lung cancer diagnosis by flagging positive cases before passing the remaining examinations to standard reporting. MATERIALS AND METHODS: A dataset of 400 CXRs including 200 difficult lung cancer cases was curated. Examinations were reviewed by three FRCR radiologists and an AI algorithm to establish performance in tumour identification. AI and radiologist labels were combined retrospectively to simulate the proposed AI triage workflow. RESULTS: When used as a standalone algorithm, AI classification was equivalent to the average radiologist performance. The best overall performances were achieved when AI was combined with radiologists, with an average reduction of missed cancers of 60%. Combination with AI also standardised the performance of radiologists. The greatest improvements were observed when common sources of errors were present, such as distracting findings. DISCUSSION: The proposed AI implementation pathway stands to reduce radiologist errors and improve clinician reporting performance. Furthermore, taking a radiologist-centric approach in the development of clinical AI holds promise for catching systematically missed lung cancers. This represents a tremendous opportunity to improve patient outcomes for lung cancer diagnosis.

Diagnosis of normal chest radiographs using an autonomous deep-learning algorithm.

AIM: To evaluate the suitability of a deep-learning (DL) algorithm for identifying normality as a rule-out test for fully automated diagnosis in frontal adult chest radiographs (CXR) in an active clinical pathway. MATERIALS AND METHODS: This multicentre study included 3,887 CXRs from four distinct NHS institutions. A convolutional neural network (CNN) was developed and trained prior to this study and was used to classify a subset of examinations with the lowest abnormality scores as high confidence normal (HCN). For each radiograph, the ground truth (GT) was established using two independent reviewers and an arbitrator in case of discrepancy. RESULTS: The DL algorithm was able to classify 15% of all examinations as HCN, with a corresponding precision of 97.7%. There were 0.33% of examinations classified incorrectly as HCN, with 84.6% of these examinations identified as borderline cases by the radiologist GT process. CONCLUSION: A DL algorithm can achieve a high level of precision as a fully automated diagnostic tool for reporting a subset of CXRs as normal. The removal of 15% of all CXRs has the potential to significantly reduce workload and focus radiology resources on more complex examinations. To optimise performance, site-specific deployment of algorithms should occur with robust feedback mechanisms for incorrect classifications.