ABS 041: Expanded Generalizability of an AI-Based Model for Pulmonary Vasculature Segmentation in Robotic Lung Surgery: From Right to Left Lower Lobe

Mohan Murari ¹ , Andres Bravo ¹ , Dr. Edward Kim ¹ , Aditya Ahuja ¹ , Danielle Birchett ¹ , Jawad Rao ² , Sara Razzaq ² , Elif Polat ² , Ritika Dinesh ¹ , Syed Abdul Khader ³ , Arian Mansur ¹ , Alexandra Nees ¹ , Omer Mescioglu ¹ , Rohin Bajaj ¹ , Dr. Sandeep Manjana ¹ , Dr. Lana Schumacher ¹

¹ Tufts University, Boston, MA
² Plaksha University, Mohali, Punjab, India
³ Harvard Medical School and Massachusetts General Hospital, Boston, MA
⁴ Massachusetts General Hospital, Boston, MA

Van Wickle (2025) Volume 1, ABS 041

Introduction: Inaccurate transections caused by misinterpretation of vascular anatomy are a significant contributor to intraoperative bleeding from pulmonary vascular injuries. Accurate segmentation of the pulmonary vasculature, particularly the pulmonary artery, is crucial for minimizing surgical risks. Building on our previous work in right lower lobe (RLL) and left upper lobe (LUL) pulmonary artery segmentation, this study evaluates the generalizability of our AI model to the left lower lobe (LLL).
We analyzed robotic lobectomy video fragments focused on the pulmonary artery and inferior pulmonary vein. Annotations from RLL lobectomy cases were used to train the XMem model, which generated segmentation masks for the pulmonary artery. The dataset was split into an 80:20 training-to-testing ratio. To test generalization, the model was evaluated on unseen LLL pulmonary artery data as well as intra-lobe inferior pulmonary vein annotations. Model performance was assessed using the Mean Intersection over Union (mIoU) metric.
The training dataset included 9 RLL pulmonary artery cases, each with 300 frames. For testing, the model was applied to 6 LLL cases, which contained fewer frames than the RLL dataset. Despite this limitation, the model achieved an average mIoU of 74.8, with scores ranging from 64.5 to 87.4. These results suggest that the AI model can generalize effectively across lobes, maintaining high segmentation accuracy.
This study highlights the potential of our AI model, initially developed for RLL and LUL pulmonary artery segmentation, to perform reliably in LLL segmentation. By enhancing intraoperative decision-making, this model could significantly improve safety in robotic lung surgery.

Methods: We trained a video object segmentation model to delineate pulmonary vascular structures in robotic lobectomy procedures. The training set consisted of nine right lower lobe (RLL) pulmonary artery cases, each manually annotated across approximately 300 frames. Segmentation masks were generated using the XMem framework, selected for its temporal consistency and performance on sequential video frames. Data was partitioned with an 80:20 split for training and internal validation. To evaluate generalizability, the model was applied to six previously unseen left lower lobe (LLL) cases, which included both pulmonary artery and inferior pulmonary vein annotations. Model performance was quantified using the mean Intersection over Union (mIoU) metric, with additional qualitative review to assess segmentation fidelity in frames with occlusions or complex vascular anatomy.

Results: We have demonstrated that an AI model trained on right lower lobe (RLL) pulmonary artery data can generalize effectively to left lower lobe (LLL) segmentation. The model achieved a mean Intersection over Union (mIoU) of 74.8% on unseen LLL cases, with individual scores ranging from 64.5% to 87.4%. Accuracy remained highest when vascular landmarks were clearly visible and declined modestly with occlusions or complex branching. These findings confirm the model’s robustness across lobar anatomy and support its potential use in intraoperative workflows to enhance pulmonary vessel identification and reduce surgical risks during robotic lung resections.

Discussion: Our results demonstrate the potential for AI-driven segmentation models to generalize across different pulmonary lobes without additional retraining. By achieving consistent performance on left lower lobe (LLL) vasculature, the model addresses a critical challenge in intraoperative navigation where anatomical variability increases surgical risk. These findings suggest that AI-assisted vascular mapping could serve as an important adjunct in robotic lung resections, enhancing both safety and precision. Future directions include expanding the dataset to incorporate a broader range of anatomical presentations and optimizing real-time performance to better integrate with intraoperative workflows, ultimately improving surgical outcomes across a wider range of thoracic procedures.