DEEP LEARNING AND ENDOSCOPIC ULTRASOUND: AUTOMATIC DETECTION AND CHARACTERIZATION OF CYSTIC AND SOLID PANCREATIC LESIONS – A MULTICENTRIC STUDY

Introduction
The diagnostic evaluation of pancreatic focal lesions remains challenging. These lesions are primarily divided into solid and cystic lesions, and their ultimate diagnosis often requires multiple exams. An accurate characterization of both cystic and solid lesions has significant implications for the definition of both treatment and follow up strategies. Endoscopic ultrasound (EUS) is involved in the diagnostic workup of these lesions. However, the diagnostic yield of this modality remains suboptimal for the differentiation within each type of lesion. Our group aimed to develop a deep learning model simultaneously providing automatic assessment for both cystic and solid pancreatic lesions.
Methods
A deep learning algorithm was generated using a total of 55,450 EUS images from 149 procedures. From these procedures, 107 (27,756 frames) were performed of the evaluation of solid lesions and 42 (27,694) for the study of cystic lesions. For solid lesions, the algorithm aimed to differentiate pancreatic ductal adenocarcinoma (PDAC) from pancreatic neuroendocrine neoplasms (PNENs) and other, less frequent, solid lesions, including a solid pseudopapillary neoplasm, a pancreatic gastrointestinal stromal tumor, a plasmacytoma, metastasis of clear cell renal cell carcinoma and accessory spleen. For cystic lesions, the algorithm was designed to provide differentiation between mucinous from non-mucinous cystic lesions. 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 curve (AUC).
Results
Figure 1 shows an example of the output provided by the CNN. The model identified PDAC with a 99.4% sensitivity, 98.6% specificity, and a global diagnostic accuracy of 99.3% (Figure 2A, AUC 0.96). The model also detected pNETs with 97.2% sensitivity and 99.8% specificity, and an accuracy of 99.4% (Figure 2B, AUC 0.99). Additionally, the model differentiated adenocarcinoma from neuroendocrine tumour with 99.4% accuracy (Figure 2C, AUC 0.99).
Regarding cystic lesions, the model differentiated mucinous from non-mucinous cystic lesions with a sensitivity of 99.1%, a specificity of 98.9% and an accuracy of 99.0% (AUC 1.00)
Conclusion
Our group developed a deep learning model capable of differentiating the main types of pancreatic solid lesions, as well as pancreatic cystic lesions, with high performance values. To our knowledge this is the first multicentric approach for the development of an AI algorithm designed to simultaneously characterize cystic and solid pancreatic lesions. The image processing time of the technology favours its clinical applicability. The development of deep learning models may help differentiate pancreatic lesions.