423 Peritoneal lesion and peri-metastatic rim classification using machine learning and image processing

Objectives/Goals: The goal of this study is to investigate the peri-metastatic rims of peritoneal lesions to determine features that predict malignancy using both imaging processing techniques and machine learning. This information will subsequently be added to our existing knowledge of peritoneal lesions to improve classification accuracy as benign or malignant. Methods/Study Population: The study population consists of 521 imaged lesions from 163 subjects with cancers of GI-origin with biopsy results as well as the clinical subject information and follow up. All images were obtained during staging laparoscopy by the senior author (TS). On the images, the central lesion as well as the surrounding peri-metastatic rim will be segmented as regions of interest (ROIs). Image processing will be used to calculate a variety of metrics for these two regions. A general estimating equation approach will be used to determine significance of these metrics compared to the dependent outcome of malignancy determined on the pathology report as the ground truth. These ROIs and significant metrics will then be used to improve the accuracy of a machine learning model to classify these lesions as benign or malignant. Results/Anticipated Results: Our previous research showed that experts performed this task at only a 52% accuracy rate (classifying lesions as malignant or benign based on imaging). A previous machine-learning model on a much smaller dataset was able to achieve by contrast an area under the curve of 0.78. We anticipate that by including a larger dataset in addition to including the peri-metastatic rim, we will be able to improve the accuracy of the the model in this task while uncovering significant biomarkers as well that can be used in future studies. Discussion/Significance of Impact: Classifying peritoneal lesions determines the correct treatment for cancer patients whether chemo-radiation, definitive surgery or palliative surgery. This project aims to develop an improved model that can perform this task using nonlabeled laparoscopic imaging with a particular focus on the diagnostic value of the peri-metastatic rim.