Development of an AI-based smart imagery framing and truthing (SIFT) system to annotate pulmonary abnormalities with corresponding boundaries based on CT images

Clinically significant pulmonary abnormalities, like tuberculosis, nodules, and chronic obstructive pulmonary disease (COPD), hinder lung function. As a non-invasive diagnostic tool, radiology can identify various pulmonary conditions. Artificial intelligence (AI) advancements, particularly deep learning-based applications, have improved the efficiency and accuracy of radiological examinations. However, creating accurate AI models is still challenging due to the requirement for extensive manual annotation. To address this, we propose Smart Imagery Framing and Truthing (SIFT) to assist annotators in labeling and delineating pulmonary abnormalities and corresponding boundaries on CT images. Integrating Mask-RCNN, the system is developed on a training dataset of 9,078 CT images with 166,724 slices and 18,367 regions of interest (ROIs) corresponding to 47 different abnormalities/diseases. The proposed system processes an independent testing set with 2,199 CT images containing 40,493 slices and 4,280 ROIs by predicting both abnormality/disease types and corresponding ROI boundary locations. Two senior radiologists (≥ 30 years of radiological experience) delineated each image of training and testing sets in a double-blinded way, and their intersection area functioned as the ground truth. Evaluation metrics include intersection over union (IOU) value calculated based on the ROI area and the ground truth, the area under the ROC curve (AUC), and sensitivity on the slice-level of abnormality label category. For ROI segmentation, the results showed that 91.5% and 36.2% abnormalities had an IOU over 0.6 and 0.7, respectively. Regarding the metric of AUC of classification, there were 97.9%, 80.9%, and 42.6% of abnormalities whose AUC values were above 0.7, 0.8 and 0.9, respectively. For the sensitivity metric, all 47 label categories exceeded 0.8, and 68.1% (32/47) exceeded 0.9. SIFT demonstrates high performance in determining abnormality/disease types with corresponding boundary locations for the ROIs, which can be used to predict training and testing sets to develop AI technologies.