AI-Driven prediction of Occult Peritoneal Metastases using Pre-operative Clinical and CT data in Gastrointestinal and Gynaecological cancers.

Introduction:

According to GLOBOCAN 2022, in the year 2022, there were 968,000 new cases of gastric cancer and close to 660,000 deaths, ranking gastric cancer as fifth in terms of both incidence and mortality worldwide, while ovarian carcinoma had 324,398 new cases and 206,839 deaths worldwide, making it the 8th most common cancer with respect to incidence and mortality among women. In India, 10056 new cases of gastric cancer were registered between 2012 and 2016, and thus it ranks second amongst the gastrointestinal cancers.  The presence of occult peritoneal metastases (PM) is currently difficult to diagnose in pre-operative radiological images, and many patients are found to have these metastases as a surprise finding during their elective surgery. This, thus, often results in abandoning a curative surgery for several reasons, such as the absence of adequate instruments or machinery for hyperthermic intra-peritoneal chemotherapy (HIPEC), or the patient might not be clinically fit for the extensive procedure, which would typically last for a minimum of 10 hours. Also, it adds to undue stress on the patient and the caretakers, especially since the patient is now termed as “metastatic” or “Stage IV” in layman’s terms. Moreover, if the presence of PM were to be detected preoperatively, the patient could have been adequately prepared for a successful cytoreductive surgery (CRS). This would then drastically improve the survival outcomes.

With the advent of artificial intelligence (AI) and deep learning (DL) in healthcare, especially in radiology, the accuracy of preoperative diagnosis has now improved in breast and brain tumors. The use of AI and DL in gastrointestinal tumors and gynecological malignancies is still in its infancy.  The DL approach is a novel method for image-based determination of complex relationships and has exhibited sophisticated performance for small feature detection and characterization. However, there is still a lack of integration of these radiological images with the intraoperative images. By having such a collaborative picture, the surgical outcomes can be improved.

Methodology:

A cross-sectional analysis of patients admitted with gastric adenocarcinoma and serous ovarian carcinoma and detected to have occult PM on staging laparoscopy or laparotomy will be performed. The patient’s clinical data and pre-operative CT imaging performed within four weeks of surgery will be obtained. The CT images will be segmentated manually for the primary and the peritoneum. The data will then be split randomly into training and test datasets using a 4:1 ratio. Using these CT images, radiomic signatures will be extracted using DL methods, and a radiomic feature set will be developed from the training cohort. The AI and DL model thus developed will be validated on the internal and external testing cohorts. Similarly, a clinical model will be developed and will be integrated with the AI/DL based CT model to produce a comprehensive model.

Conclusion:

The AI/DL model thus developed will be able to predict the occult peritoneal metastases on pre-operative imaging, avoiding unnecessary surgeries and/or providing personalised surgical planning.