Automatic differentiation of solid pancreatic lesions in Endoscopic Ultrasound using – a multicentre study

Aims Solid pancreatic lesions present a diagnostic challenge, often requiring multiple exams in the typical diagnostic workup. Pancreatic adenocarcinoma is the most common solid lesion, associated with a poor prognosis due to late-stage diagnosis. Endoscopic ultrasound (EUS) is typically involved in the diagnostic workup of these lesions. However, the diagnostic yield of this modality remains suboptimal, and it is often difficult to differentiate adenocarcinoma from other solid lesions. Our group aimed to develop a deep learning model for the automatic differentiation of solid pancreatic lesions, focusing on the identification of the most common pancreatic neoplasms: pancreatic ductal adenocarcinoma (PDAC) and pancreatic neuroendocrine tumors (pNETs).

Methods A total of 27,756 images from 107 EUS exams conducted in two specialized centres were used for the development of the convolutional neural network. The dataset comprised 22,710 images of PDAC, 3,886 images of pNETs, and 1,160 images of other solid pancreatic findings, including a solid pseudopapillary neoplasm, a pancreatic gastrointestinal stromal tumor, a plasmacytoma, metastasis of clear cell renal cell carcinoma and and accessory spleen. The training dataset included approximately 90% of the total images, while the testing dataset, used to evaluate the model, consisted of the remaining 10%. The model was evaluated through its sensitivity, specificity, positive and negative predictive values, accuracy, and area under the precision-recall curve (AUC-PR).

Results The model identified PDAC with a 99.4% sensitivity, 98.6% specificity, a positive predictive value of 99.7%, and a negative predictive value of 97.4%, achieving a global diagnostic accuracy of 99.3%. The model also detected pNETs with 97.2% sensitivity and 99.8% specificity, with positive and negative values of 98.5% and 99.5%, respectively, and a global accuracy of 99.4%. Additionally, the model differentiated adenocarcinoma from neuroendocrine tumour with 99.4% accuracy. The convolutional neural network had an AUC-PR of 0.85 for the identification of pancreatic adenocarcinoma, with an AUC-PR of 1.00 for the identification of neuroendocrine tumours and differentiation between adenocarcinoma and neuroendocrine tumours.

Conclusions Our group developed a deep learning model capable of differentiating PDAC from pNETs and other solid pancreatic lesions with accuracy. The image processing time of the technology favours its clinical applicability. The development of deep learning models may help differentiate solid pancreatic lesions, achieving a more accurate diagnosis of pancreatic adenocarcinoma.

Publikationsverlauf

Artikel online veröffentlicht:
15. April 2024