New England Surgical Society
NESS Home NESS Home Past & Future Meetings Past & Future Meetings

Back to 2023 Abstracts


Development of a Deep Learning System for Prediction of Chemotherapy Resistance from Staging Laparoscopy Images
Thomas Schnelldorfer*1, Janil Castro1, Francis W. Nugent2, Georgios Georgalis3
1Division of Surgical Oncology, Tufts Medical Center, Boston, MA; 2Division of Hematology / Oncology, Tufts Medical Center, Boston, MA; 3Data Intensive Studies Center, Tufts University, Medford, MA


Objective: For patients with metastatic gastrointestinal cancers, chemotherapy resistance is a crucial measure that if known in advance could be used to make patient-centric and targeted treatment decisions. Yet, this measure currently cannot be captured in advance. While the degree of chemotherapy resistance can vary tremendously between patients, so can the optical appearance of cancer metastases. The metastasis" optical appearance, as a summary of its molecular makeup, can be measured during operative staging and might provide guidance for subsequent treatment decisions.
Design: Retrospective observational study for development of a convolutional deep neural network to predict chemotherapy resistance from standard laparoscopy images depicting peritoneal metastases.
Setting: Single tertiary care hospital.
Patients: Twenty-seven adult patients who underwent staging laparoscopy for non-colon gastrointestinal adenocarcinoma (14 gastric, 10 pancreatic, 3 other) with biopsy-proven peritoneal metastases between 2014 and 2021 and received cytotoxic chemotherapy.
Main Outcome Measure: Accuracy of a deep learning system to correctly classify chemotherapy resistance on laparoscopy images of peritoneal metastases.
Results: Each patients" cancer-specific survival was used to determine if a patient was considered chemotherapy sensitive (i.e. observed survival was greater than expected median) or chemotherapy resistant (i.e. observed survival was less than expected median) when controlled for the amount of chemotherapy received. Of 27 patients, 15 were assigned to the chemotherapy sensitive group and 12 to the chemotherapy resistant group. There was no significant difference in age, performance status, cancer type, peritoneal cancer index, and amount and timing of chemotherapy given between both groups (all p>0.293). The study cohort provided 800 image patches of biopsy-confirmed peritoneal metastasis (10 per metastasis from different angles and distances). When testing a wide range of convolutional neural networks, overfitting was a consistent problem. Therefore, the architecture of the deep learning system included image augmentations and dropout layers. The system was evaluated using a random 80:20 train-test split on lesion level. With a prediction threshold of 0.5, the system provided an accuracy of 0.61 (95% CI, 0.53-0.70) in predicting the correct group. This corresponded to a sensitivity of 0.67 and a specificity of 0.55. In addition, the system was tested on 1266 image patches depicting all visible non-biopsied peritoneal lesions within the study cohort. This dataset demonstrated an accuracy of 0.55 (95% CI, 0.52-0.58).
Conclusion: In this preliminary study, the use of a deep learning surgical guidance system designed to intra-operatively identify chemotherapy resistance from peritoneal metastases demonstrated its technical feasibility. Further development of the prototype and subsequent validation in a multi-institutional clinical study is needed.


Back to 2023 Abstracts