# Development and Validation of a CT Radiomics-Deep Learning Model for Predicting Surgical Difficulty in Pancreatic and Periampullary Tumors

## TL;DR

This study creates a model using CT scans to predict the difficulty of a specific cancer surgery, helping doctors prepare better.

## Contribution

A novel CT radiomics-deep learning model is developed for predicting surgical difficulty in pancreatic and periampullary tumors.

## Key findings

- The combined model achieved a test set AUC of 0.848 and high sensitivity of 0.850 for identifying difficult surgical cases.
- The model outperformed standalone radiomics models with a testing AUC of 0.848 compared to 0.754 and 0.816 for other models.

## Abstract

In this retrospective study, we developed and validated an integrated CT radiomics-deep learning model (RDLM) for preoperative prediction of LPD surgical difficulty. The model combines hand-crafted radiomics features (intratumoral and peritumoral) and deep learning-derived features, achieving a test set AUC of 0.848 and high sensitivity (0.850) for identifying difficult cases. Key strengths include non-invasiveness, robust calibration, and clinical net benefit. Contextualized within the field, this model addresses the unmet need for preoperative risk stratification in LPD, complementing existing surgeon-dependent assessments.

Background: Pancreatic and periampullary cancers are common tumors of the digestive tract. As a radical surgical approach, laparoscopic pancreaticoduodenectomy requires crucial preoperative assessment of its surgical difficulty. Materials and methods: A retrospective cohort of 150 patients who underwent LPD between June 2019 and June 2023 was enrolled. The criteria for defining the difficult group were identified as unplanned conversion to open procedure, intraoperative blood loss, and operative time. Participants were randomly allocated to a training set (n = 105) or a testing set (n = 45) in a 7:3 ratio. Hand-crafted radiomics (HCR) features and deep learning-derived radiomics (DLR) features were extracted from portal venous phase CT images, focusing on gross tumor volume and gross peri-tumor volume. A hybrid prediction model was developed using a support vector machine algorithm, with performance evaluated through receiver operating characteristic analysis, calibration curves, and decision curve analysis (DCA). Results: The combined model demonstrated significantly superior discriminative ability, achieving an area under the curve (AUC) of 0.942 (95% CI: 0.893–0.992) in the training set and 0.848 (95% CI: 0.738–0.958) in the testing set. This performance exceeded both the standalone HCR model (testing AUC = 0.754) and the DLR model (testing AUC = 0.816). DCA further confirmed the clinical utility of the combined model, showing the highest net benefit across threshold probabilities exceeding 20%. Conclusions: The novel integrated model combining hand-crafted and deep learning-derived radiomics features enables effective prediction of surgical difficulty in laparoscopic pancreaticoduodenectomy.