Improving the Generalizability and Performance of an Ultrasound Deep Learning Model Using Limited Multicenter Data for Lung Sliding Artifact Identification.

Deep learning (DL) models for medical image classification frequently struggle to generalize to data from outside institutions. Additional clinical data are also rarely collected to comprehensively assess and understand model performance amongst subgroups. Following the development of a single-center model to identify the lung sliding artifact on lung ultrasound (LUS), we pursued a validation strategy using external LUS data. As annotated LUS data are relatively scarce-compared to other medical imaging data-we adopted a novel technique to optimize the use of limited external data to improve model generalizability. Externally acquired LUS data from three tertiary care centers, totaling 641 clips from 238 patients, were used to assess the baseline generalizability of our lung sliding model. We then employed our novel Threshold-Aware Accumulative Fine-Tuning (TAAFT) method to fine-tune the baseline model and determine the minimum amount of data required to achieve predefined performance goals. A subgroup analysis was also performed and Grad-CAM++ explanations were examined. The final model was fine-tuned on one-third of the external dataset to achieve 0.917 sensitivity, 0.817 specificity, and 0.920 area under the receiver operator characteristic curve (AUC) on the external validation dataset, exceeding our predefined performance goals. Subgroup analyses identified LUS characteristics that most greatly challenged the model's performance. Grad-CAM++ saliency maps highlighted clinically relevant regions on M-mode images. We report a multicenter study that exploits limited available external data to improve the generalizability and performance of our lung sliding model while identifying poorly performing subgroups to inform future iterative improvements. This approach may contribute to efficiencies for DL researchers working with smaller quantities of external validation data.

Anatomic evaluation of Pancoast tumors using three-dimensional models for surgical strategy development.

OBJECTIVE: Pancoast tumor resection planning requires precise interpretation of 2-dimensional images. We hypothesized that patient-specific 3-dimensional reconstructions, providing intuitive views of anatomy, would enable superior anatomic assessment. METHODS: Cross-sectional images from 9 patients with representative Pancoast tumors, selected from an institutional database, were randomly assigned to presentation as 2-dimensional images, 3-dimensional virtual reconstruction, or 3-dimensional physical reconstruction. Thoracic surgeons (n = 15) completed questionnaires on the tumor extent and a zone-based algorithmic surgical approach for each patient. Responses were compared with surgical pathology, documented surgical approach, and the optimal "zone-specific" approach. A 5-point Likert scale assessed participants' opinions regarding data presentation and potential benefits of patient-specific 3-dimensional models. RESULTS: Identification of tumor invasion of segmented neurovascular structures was more accurate with 3-dimensional physical reconstruction (2-dimensional 65.56%, 3-dimensional virtual reconstruction 58.52%, 3-dimensional physical reconstruction 87.50%, P < .001); there was no difference for unsegmented structures. Classification of assessed zonal invasion was better with 3-dimensional physical reconstruction (2-dimensional 67.41%, 3-dimensional virtual reconstruction 77.04%, 3-dimensional physical reconstruction 86.67%; P = .001). However, selected surgical approaches were often discordant from documented (2-dimensional 23.81%, 3-dimensional virtual reconstruction 42.86%, 3-dimensional physical reconstruction 45.24%, P = .084) and "zone-specific" approaches (2-dimensional 33.33%, 3-dimensional virtual reconstruction 42.86%, 3-dimensional physical reconstruction 45.24%, P = .501). All surgeons agreed that 3-dimensional virtual reconstruction and 3-dimensional physical reconstruction benefit surgical planning. Most surgeons (14/15) agreed that 3-dimensional virtual reconstruction and 3-dimensional physical reconstruction would facilitate patient and interdisciplinary communication. Finally, most surgeons (14/15) agreed that 3-dimensional virtual reconstruction and 3-dimensional physical reconstruction's benefits outweighed potential delays in care for model construction. CONCLUSIONS: Although a consistent effect on surgical strategy was not identified, patient-specific 3-dimensional Pancoast tumor models provided accurate and user-friendly overviews of critical thoracic structures with perceived benefits for surgeons' clinical practices.

Augmented Reality and Intraoperative Navigation in Sinonasal Malignancies: A Preclinical Study.

OBJECTIVE: To report the first use of a novel projected augmented reality (AR) system in open sinonasal tumor resections in preclinical models and to compare the AR approach with an advanced intraoperative navigation (IN) system. METHODS: Four tumor models were created. Five head and neck surgeons participated in the study performing virtual osteotomies. Unguided, AR, IN, and AR + IN simulations were performed. Statistical comparisons between approaches were obtained. Intratumoral cut rate was the main outcome. The groups were also compared in terms of percentage of intratumoral, close, adequate, and excessive distances from the tumor. Information on a wearable gaze tracker headset and NASA Task Load Index questionnaire results were analyzed as well. RESULTS: A total of 335 cuts were simulated. Intratumoral cuts were observed in 20.7%, 9.4%, 1.2,% and 0% of the unguided, AR, IN, and AR + IN simulations, respectively (p < 0.0001). The AR was superior than the unguided approach in univariate and multivariate models. The percentage of time looking at the screen during the procedures was 55.5% for the unguided approaches and 0%, 78.5%, and 61.8% in AR, IN, and AR + IN, respectively (p < 0.001). The combined approach significantly reduced the screen time compared with the IN procedure alone. CONCLUSION: We reported the use of a novel AR system for oncological resections in open sinonasal approaches, with improved margin delineation compared with unguided techniques. AR improved the gaze-toggling drawback of IN. Further refinements of the AR system are needed before translating our experience to clinical practice.