Back to 2025 Abstracts
Low-Compute Techniques for Peritoneal Lesion Detection During Staging Laparoscopy
Isaac Gendelman
*1, Janil Castro
1, Sebastian Schnelldorfer
2, Liping Liu
2, Thomas Schnelldorfer
11Surgical Imaging Lab, Tufts Medical Center, Boston, MA; 2Computer Science, Tufts University, Medford, MA
BACKGROUND: Identification of every peritoneal lesion during staging laparoscopy is crucial for the surgical care of various malignancies. Prior studies suggested a significant rate of peritoneal surface metastases being missed by oncologic surgeons. As a result, there have been advances using artificial intelligence (AI) models for peritoneal lesion detection and classification. The current study assesses whether less compute-heavy image processing techniques can prove useful in augmenting AI approaches to detect peritoneal lesions while providing significantly reduced computational cost. STUDY DESIGN: Staging laparoscopy images were obtained from an image library of patients with non-colon gastrointestinal adenocarcinoma. An algorithm was developed that combines a Gabor texture filter, a Canny edge-detection filter, three-channel color filters and blob smoothing to segment unlabeled images. The results were compared to images where the region of interest (ROI) was manually segmented, providing the study’s ground truth. The parameters of the algorithm were tuned on a smaller subset of data to achieve adequate accuracy before testing on the entire cohort. Results were then compared to our previously published AI algorithm. RESULTS: The developed algorithm was tested on a cohort of 1414 images capturing 5943 lesions from 163 patients undergoing staging laparoscopy. These patients had a mean peritoneal cancer index of 2.5 +/- 4.7 and 57 had metastatic disease while 106 did not. The mean accuracy of the algorithm was 0.72 +/- 0.19 with a mean recall of 0.59 +/- 0.39 and a mean precision of 0.05 +/- 0.11. The mean total segmented area of the image (or predicted positive rate, PPR) was 0.28 +/- 0.19. Thus the algorithm was able to exclude 72% of the image area. The precision-recall area under the curve (PR-AUC) was 0.08 compared to 0.69 for the AI system. CONCLUSIONS: Low-compute approaches such as image filtering achieve reasonable accuracy and recall but low precision. This method is most useful in excluding a majority of the image as negative at a low computational cost and ideally should be combined with an AI model for classification of the predicted ROI as part of a system for automated identification of peritoneal surface metastases.
Back to 2025 Abstracts
