For patients who undergo operative resections for gastrointestinal cancers, treatment selection fundamentally relies on the result of intra-operative assessment of the extent of the underlying cancer (i.e. staging). Specifically, the absence or presence of distant metastases dictates the role of operative treatment, chemotherapy, and radiation. However, the accuracy of operative staging (i.e. staging laparoscopy) is limited resulting in ?under-staging? in up to 30% of these patients adversely affecting their cancer treatment. While operative ?under-staging? is thought to equally affect many other malignancies, the cause is believed to arise from the inability of a conventional operative exam to reliably differentiate benign from metastatic lesions. Recent results demonstrated that expert surgeons on average misidentify 3619% of grossly visible metastases questioning the accuracy of a human examiner. Our long-term goal is to significantly improve the accuracy of operative staging laparoscopy in patients with gastrointestinal cancers by enhancing its capability to detect metastases through means of machine learning. To achieve this goal, we will use existing videos from staging laparoscopies and abstract images of peritoneal lesions that underwent biopsy (i.e. ground truth) as part of routine care (Aim 1). These images will then be used for the development of an automated classification system. The first step of developing the classification system involves training of a deep neural network with weak supervision that will allow for automated segmentation of lesions from their surrounding background (Aim 2). The second step will extract feature vectors from the lesions segmented in Aim 2 providing information for classification. The feature vectors will be extracted by two parallel processes: unsupervised deep learning and extraction of expert-selected features. The resulting feature vectors will be used to train a model allowing the classification (benign vs. metastasis) of any peritoneal lesion (Aim 3). The results of this study are expected to provide material for future improvements / modifications of the proposed deep learning classification system as well as the foundation for future development of an automated surgical guidance system designed to help surgeons reliably identify metastases. Relevance: This study will establish a robust, yet simple method to improve the staging accuracy of standard laparoscopy via the detection of peritoneal metastases otherwise missed by human examiners. This will significantly improve cancer care through better treatment allocation. Further, it is expected that the detection of currently missed metastases will have a major impact on staging and treatment algorithms for a variety of cancers.
During operations to treat gastrointestinal cancers, disease spread to other sites (i.e. metastases) is not recognized in a significant proportion of patients adversely affecting their cancer care. The proposed study will utilize artificial intelligence computer algorithms that will allow for automated identification and classification of such metastases. The results are expected to provide the foundation for future development of an automated surgical guidance system meant to enhance operative detection of metastases.
