SRS60 - Evaluation of deep learning-based image analysis for diagnostic and surgical planning using the rsna pneumonia detection dataset

Abstract Introduction AI-powered image analysis provides rapid, reproducible diagnoses that are critical during surgical planning and intraoperative decision-making. The purpose of this study was to assess the diagnostic performance of a deep learning algorithm trained on a large, annotated, open-access imaging dataset and to measure its influence on surgical workflow efficiency and team confidence. Methods We analysed the RSNA Pneumonia Detection Challenge dataset, which contains 26 684 de-identified chest radiographs, to train and validate a convolutional neural network (CNN) for detection of pneumonia. The dataset was split with 85 percent of images for training and 15 percent for validation. Model performance was evaluated by accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC). Results The AI model achieved an AUC of 0.90 (95% c.i.: 0.89–0.91), sensitivity of 0.87, and specificity of 0.83 for detecting radiographic pneumonia. Use of the AI assistant led to a 21 percent reduction in interpretation time for surgical teams reviewing preoperative images. In simulated scenarios, radiologists and surgeons reported higher confidence in preoperative decision-making when aided by AI outputs. Discussion Deep learning-based analysis of large, publicly accessible datasets such as RSNA’s provides reliable diagnostic support and actionable workflow gains for surgeons. The observed improvements in accuracy and efficiency underscore the value of integrating AI into surgical pathways. Continued research is warranted to prospectively validate clinical impact across a range of surgical indications using robust, real-world datasets.