Liquid-based cytological diagnosis of pancreatic neuroendocrine tumors using hyperspectral imaging and deep learning

The incidence of pancreatic neuroendocrine tumors (PanNETs), although uncommon, has recently increased, and they are almost always diagnosed in the late stages due to their indolent clinical manifestations. This study developed a method that combines hyperspectral imaging (HSI) technology and a convolutional neural network (CNN) to conduct a cytological diagnosis of PanNETs. We acquired hyperspectral information from the nuclei of PanNETs and benign cells derived from liquid-based cytology (LBC) of pancreatic endoscopic ultrasound-guided fine-needle aspiration/biopsy specimens. CNN model was trained to distinguish among different cell types based on spectral and spatial differences compared with conventional “red, green, and blue (RGB)” images. For cell classification, the CNN system identified hyperspectral images containing PanNETs and benign cells with areas under the curve (AUCs) of 0.9981 and 0.9815 in datasets from two different time points, respectively, showing superior performance compared to the conventional RGB group, with corresponding AUCs of 0.9716 and 0.9550. Higher accuracy was achieved for the HSI group than for the RGB group in both test datasets (94.92 ​% versus 89.85 ​% and 93.19 ​% versus 80.63 ​%, respectively). Our results revealed that HSI-based cytological diagnosis using a CNN could provide superior classification performance for PanNETs compared with conventional RGB images. With further validation, this innovative technique can be utilized as an alternative to traditional cytological diagnosis for higher efficiency, thus reducing the workload of daily clinical practice.