Objectives/Goals: Accurate detection of peritoneal metastases during staging laparoscopy is critical for treatment planning, yet even skilled surgeons miss lesions. Thus, an optimized low-compute image-processing pipeline that maximizes recall while filtering non-lesion tissue can assist with this task. Methods/Study Population: The study population consists of 1414 images containing 5943 annotated lesions from 163 subjects undergoing staging laparoscopy for GI-origin cancer. All images were obtained during staging laparoscopy by the senior author (TS). From the original images, five binary masks were produced using red, green, blue, Canny, and Gabor filters. These were combined into a composite image, thresholded, and binarized after smoothing into a final prediction. Given the complex 16-dimensional space, a genetic fitness algorithm was used to optimize the final set of parameters to maximize recall while also penalizing coverage. Results/Anticipated Results: Preliminary data show that the algorithm has a mean recall of 0.90 ± 0.18 and a mean predicted positive rate of 0.48 ± 0.20. Thus, the algorithm is able to exclude 52% of the image area while capturing 90% of lesion pixels. We plan to further validate these results and maximize this pipeline to demonstrate that simple tools can yield useful results. Discussion/Significance of Impact: This project aims to develop a computationally simple image-processing pipeline that can achieve high recall while minimizing predicted positive coverage. This will be useful to highlight where lesions are most likely to exist and may be combined in the future with classification models.
Gendelman et al. (Wed,) studied this question.