Taha, Abtahir Rahim

Abstract RATIONALE: Research since 2017 has highlighted artificial intelligence (AI)-based technologies to assist in detecting and/or classifying lung nodules on Computed Tomography (CT) scans. However, few have combined detection and classification tasks, and most have required extensive hands-on image manipulation to process the cases, which is both costly and time-consuming. Here, we present a closed loop, fully automated software package, Bronchosolve, which handles cases from scan to result without manual intervention. METHODS: The software package combined pre-processing, analysis, and result generation system including a deep learning convolutional neural network (CNN) for assessment of suspicious pulmonary nodules. Model inputs were full CT cases formatted in Digital Imaging and Communications in Medicine (DICOM), without demographics or clinical variables. Automated steps included 1) automatic selection of the optimal CT series for analysis; 2) normalization and pre-processing of the series; 3) AI-based analysis of the images to both detect and classify suspicious lung lesions; and 4) generation of analysis report. The system was trained using a multi-center dataset of over 2,358 cases of suspicious lung nodules, including cancers and non-cancers. A United States-based, multi-site cohort (n=184) was used for initial algorithm validation where positive cases were defined as biopsy-proven cancer within 1 year, and negatives as negative biopsy or 2+ year negative follow-up. RESULTS: In the validation set, the system achieved a technical completion rate of 100%, automatically for all cases. Median patient age was 62.5 years (interquartile range 58.5 to 66.5 years), 45% were former smokers, 55% were current smokers, and 40% were female. The AI cancer risk (CA-risk) model achieved an area under the receiver-operating-characteristic curve (AUC) of 0.828 [CI: 0.773 – 0.883] in differentiating cancer from non-cancer in cases with known suspicious pulmonary nodules, as confirmed by biopsy, surgery, and/or follow-up. These results were statistically superior to Lung-RADS pAUC of 0.669 [CI: 0.601 – 0.737] and trended better than Brock model AUC of 0.783 [CI: 0.723 – 0.842], Figure 1. At the Youden operating point, model sensitivity was 73.1% [CI: 66.7 – 79.5%], and specificity was 82.5% [CI: 77.0 – 88.0%]. The model's performance was consistent across various CT scanner manufacturers and tested slice thickness. CONCLUSIONS: The closed loop, automated software system, Bronchosolve, consistently processed and analyzed a wide range of CT scans, providing accurate assessment in risk-classifying suspicious lung nodules. This approach has the potential to improve non-invasive assessment of lung nodules eliminating the need for additional user training and time.