Deep learning via pathology-image integration for margin prediction after pancreatoduodenectomy

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Positive surgical margins (R1) after pancreatoduodenectomy for pancreatic ductal adenocarcinoma (PDAC) are frequent and linked to poorer survival.
• Preoperative imaging often fails to detect microscopic margin involvement, especially near the superior mesenteric artery.
• Accurate prediction of margin status is crucial for surgical planning and patient outcomes.

Purpose of the Study:

• To evaluate the efficacy of deep learning algorithms applied to preoperative CT scans for predicting pathological margin status in PDAC.
• To assess the impact of image preprocessing techniques on the performance of deep learning models.

Main Methods:

• A cohort of 100 patients undergoing pancreatoduodenectomy for PDAC was analyzed.
• Two convolutional neural network models were developed using an attention-based multiple instance learning approach on triphasic CT scans.
• Model 2 incorporated noise-reduction techniques, including selective slice extraction and normalization, to enhance performance.

Main Results:

• The enhanced deep learning model (Model 2) achieved an Area Under the Curve (AUC) of 0.78, outperforming the baseline model (AUC 0.69).
• Model 2 demonstrated high sensitivity (88.2%) in identifying margin-positive cases (R1/R2).
• Noise-reduction preprocessing significantly improved the predictive performance of the deep learning model.

Conclusions:

• Deep learning applied to preoperative CT imaging is a feasible tool for predicting pathological margin status after pancreatoduodenectomy.
• Image noise-reduction techniques enhance the accuracy of AI models for margin prediction.
• This AI-driven approach can aid in preoperative risk stratification and surgical planning for PDAC patients.