BACKGROUND Real-time anatomical recognition during robot-assisted surgery has the potential to enrich intraoperative decision-making. We made the first clinical test of an artificial intelligence (AI)-based algorithm designed to identify key anatomical structures during robot-assisted retroperitoneal renal and adrenal procedures using the daVinci single-port (SP) platform. METHODS A deep learning algorithm, trained on annotated surgical video frames and retrained with an expanded dataset, was implemented to identify six anatomical structures: the psoas muscle, ureter, kidney, renal artery, renal vein, and inferior vena cava (IVC). The system's performance was prospectively evaluated in 15 patients operated on from April 2025 to July 2025, comparing the timing of recognition (Δ-time) with that of the operating surgeon, and recognition patterns (first, simultaneous, or exclusive identification) were recorded. The impact of patient body mass index (BMI) and history of previous abdominal surgery on recognition time were evaluated. RESULTS The surgeon was most frequently the first to recognize anatomical structures, including the psoas muscle (86.7%), renal artery (86.7%), and kidney (73.3%). The AI system achieved simultaneous recognition in a subset of cases but was rarely the first recognizer (psoas muscle 6.7%, renal vein 13.3%). No structures were identified exclusively by the AI, while several were recognized solely by the surgeon. Median Δ-time values ranged from 12.0 seconds (IVC) to 85.0 seconds (renal vein), with statistically significant differences for most structures (p < 0.01), except IVC (p = 0.100). In our analysis of recognition time, both BMI > 30 (p = 0.037) and a history of previous abdominal surgery (p = 0.019) were significantly associated with longer AI recognition time for the renal artery. CONCLUSIONS The AI algorithm demonstrated the capability to recognize key anatomical structures in real time, although the surgeon consistently outperformed it in terms of recognition timing. This tool shows promise as a supportive system for intraoperative guidance, particularly in enhancing situational awareness, with future improvements needed to reach autonomous-level recognition accuracy.
BACKGROUND Single-incision laparoscopic cholecystectomy (SILC) is technically more challenging and has a steeper learning curve compared to conventional multi-port laparoscopic cholecystectomy. A deep learning-based system for surgical phase recognition and prediction in SILC can facilitate surgical quality assessment, enhance training efficiency, and support safer procedural execution. In this study, we aimed to develop a multicenter surgical phase recognition and prediction system for SILC. METHODS This study included 148 SILC videos collected from two medical centers. Videos were labeled by surgeons, and a deep learning model was developed based on 122 videos. The performance of the model was tested in an additional 26 videos by comparing it with the annotated ground truth of the surgeon. Deep learning models were trained to identify SILC phases. The performance of models was measured. For surgical phase classification, the following metrics were used: accuracy, precision, recall, and Jaccard Index. For surgical phase prediction, a variant of mean absolute error (MAE) was adopted. RESULTS The analyzed SILC videos had a mean operative duration of 14.36 min, with considerable variability across cases. The Trans-SVNet model achieved strong performance in surgical phase classification, with overall accuracy, precision, and recall of 0.933, 0.939, and 0.939, respectively. In terms of phase transition prediction, the model yielded an overall inMAE of 37 s, eMAE of 34 s, and pMAE of 50 s. Notably, increasing the number of training cases led to significant improvements in model performance. CONCLUSIONS We successfully developed the Trans-SVNet model, which enabled automatic classification and temporal prediction of surgical phases in full-length SILC videos. With continued refinements, artificial intelligence could be utilized in huge data surgery analysis to achieve clinically relevant future applications.
Despite the widespread use of minimally invasive surgery, bile duct injury (BDI) remains a significant complication of laparoscopic cholecystectomy (LC). Accurate identification of the common bile duct (CBD) and common hepatic duct (CHD) is critical for preventing BDI. This study aimed to develop and evaluate a deep learning model for segmenting the “CBD zone,” a semantically defined high-risk region within the hepatoduodenal ligament encompassing the CHD, CBD, and adjacent arterial structures, to assist with anatomical recognition during LC. Surgical videos from 100 patients who underwent elective LC, with a total recorded duration of less than 20 minutes, were retrospectively analyzed. Three board-certified hepatobiliary surgeons annotated the CBD zone and gallbladder. A total of 1045 frames were extracted and divided into training, tuning, and internal test sets. Each case was graded according to the Parkland Grading System to stratify cholecystitis severity. Three convolutional neural networks – U-Net, LRASPP-MobileNetV3, and DeepLabv3-ResNet101 – were trained. Segmentation performance was evaluated using the Dice score and Intersection over Union (IoU). Simulation-based expert assessment was performed on internal and external datasets using pass/fail ratings. DeepLabv3 demonstrated the best segmentation performance, with Dice scores of 0.910 for the CBD zone and 0.903 for the gallbladder. The mean IoU was 0.857 for the CBD zone and 0.847 for the gallbladder. In simulation-based expert assessment, the mean pass rate across three internal experiments was 96.0%. In the external dataset, the mean pass rate was 75.8%. We developed an AI-based segmentation model for identifying the CBD zone and gallbladder during LC. The model demonstrated strong segmentation performance and preliminary feasibility in external validation. Further prospective and multicenter studies are required to determine its practical utility in intraoperative settings.
OBJECTIVES To evaluate the feasibility and accuracy of an artificial intelligence (AI) model to assist surgeons through automated real-time detection and segmentation of key anatomical structures during robot-assisted single-port transvesical enucleation of the prostate (STEP). PATIENTS AND METHODS This retrospective single-centre study utilised surgical videos from patients undergoing single-port robot-assisted transvesical prostate enucleation performed by a single expert surgeon. Selected frames extracted from these surgical videos were manually annotated to identify anatomical landmarks during the key step of the procedure. The structures annotated were the bladder neck, prostatic adenoma, and the peripheral zone. A convolutional neural network based on the You Only Look Once version 11 architecture was trained using these annotated frames. Model performance was quantitatively assessed through recall, precision, F1-score, Intersection over Union, Dice Similarity Coefficient, and mean average precision (AP). Real-time performance was assessed qualitatively through visual inspection and confirmed quantitatively by measuring frame inference speeds. RESULTS The study included 611 annotated frames derived from 37 surgical videos. The model demonstrated strong detection performance, with class-specific F1-scores of 0.88 (adenoma), 0.69 (bladder neck), and 0.70 (peripheral zone). Segmentation accuracy, measured by Dice Similarity Coefficient, resulted in adenoma: 0.86; bladder neck: 0.83; peripheral zone: 0.82. Mean AP across all anatomical classes was 0.43. The model consistently operated at >60 frames/s, confirming real-time applicability without perceptible lag. CONCLUSIONS This pilot study demonstrates the potential utility of AI to provide intraoperative anatomical guidance during STEP, establishing a foundation for future clinical integration and performance optimization.
BACKGROUND Curative treatment of resectable esophageal cancer comprises neoadjuvant chemoradiotherapy and esophagectomy. Robot-assisted minimally invasive esophagectomy (RAMIE) is the preferred technique; however, learning RAMIE is challenging due to the complex chest anatomy, patient positioning and zoomed-in camera view. Computer-aided anatomy recognition holds promise for improving surgical navigation. This study aims to develop real-time anatomy recognition and surgical phase recognition algorithms for the thoracic part of RAMIE using deep learning and to understand the challenges of current state-of-the-art algorithms. METHODS A retrospective single-center cohort study was conducted on prospectively collected RAMIE videos at University Medical Center Utrecht, The Netherlands. Two datasets were created: an anatomy segmentation dataset with 1504 frames from 53 videos, annotated for eight anatomical classes and four surgical instruments; and a surgical phase dataset, with 38 videos labeled with thirteen distinct phases. Several deep learning models were trained and tested using both datasets. RESULTS The SegNeXt model achieved the most accurate segmentations with an overall overlap score of 0.72 for all classes, of which the surgical instruments were best detectable. The best surgical phase recognition model achieved an overall accuracy of 82.8% and revealed loss of accuracy during phase transitions. Class imbalance affected both datasets which led to less frequent appearing classes performing lowest. CONCLUSIONS This study identified suitable models for computer-aided, real-time anatomy and surgical tool segmentation and surgical phase recognition for the thoracic part of RAMIE. Acceptable performance was achieved while both datasets are highly complex. Performance is expected to improve by further expanding and diversifying the datasets.
While the results are promising for a novel application domain, the model's failure on clinically critical minority classes (Bilateral Blockage, Bilateral Patency) means it is not yet suitable for unsupervised clinical use.
We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.
