Differentiating cytology of pancreatic ductal adenocarcinoma and pancreatic neuroendocrine tumors by EUS-FNA through hyperspectral imaging technology combined with artificial intelligence.

[BACKGROUND] Pancreatic cancer is a common and lethal malignancy, with the two primary subtypes being pancreatic ductal adenocarcinoma (PDAC) and pancreatic neuroendocrine tumors (pNET). Accurate diagnosis and effective treatment are crucial. Hyperspectral imaging (HSI) is a novel optical diagnostic technology that can capture spectral features inaccessible by traditional imaging techniques. With the aid of artificial intelligence (AI), HSI can provide richer information.

[OBJECTIVES] This study aims to develop a convolutional neural network (CNN) based on HSI to assist in the diagnosis of liquid-based cytology (LBC) specimens of PDAC and pNET obtained by endoscopic ultrasound-guided fine-needle aspiration (EUS-FNA).

[DESIGN] We designed a deep learning model using HSI data to differentiate between PDAC and pNET specimens. The CNN model was developed and evaluated using a dataset of LBC slides.

[METHODS] During the EUS-FNA procedure, we prepared LBC slides of PDAC and pNET specimens. These slides were scanned using HSI technology to acquire both spectral and spatial information. We employed a modified ResNet18 model to analyze this information and perform classifications. In addition, we used attribute-guided factorization visualization (AGF-visualization) to visualize the CNN's decision-making process.

[RESULTS] Based on samples from 59 patients, 2014 HSI images were acquired. The spectral curves of PDAC and pNET cells exhibited recognizable differences in the wavelength range of 520-600 nm. Our modified ResNet18 model processes images at approximately 9 images/s and achieves a sensitivity of 90.80%, a specificity of 94.68%, and an accuracy rate of 92.82% (area under the receiver operating characteristic curve = 0.9721). AGF-visualization confirmed that our CNN model classifies based on the features of the tumor cell nucleus.

[CONCLUSION] Our HSI-CNN model accurately differentiates PDAC and pNET in EUS-FNA specimens, aiding pathologists in diagnosis and reducing their workload.