Pancreatic Cancer (PC) is one of the most aggressive cancers, constituting the seventh leading cause of cancer-related death globally in 2020. Usually, the asymptomatic response of PC causes the delayed diagnosis of the disease. Diagnosis of PC usually includes ultrasonography (US), computed tomography (CT), magnetic resonance (MRI), and endoscopic ultrasound (EUS). Although EUS is the diagnostic method with the highest sensitivity reported, the procedure is highly operator-dependent. A gastroenterologist requires more than 150 supervised procedures to interpret the anatomy blurred by several noise sources. Therefore, a second reader may be desirable to support the procedure and assist the training process in a gastroenterology service. Some computational strategies have been developed to detect PC in EUS images, but those methods are semi-automatic in practice and very susceptible to noise. Hence, the main contribution of this work is an automatic strategy to detect PC in complete video sequences of EUS procedures. The proposed methodology describes the mixture of echo patterns using the Speeded-Up Robust Features (SURF) method. A set of interest points are defined and described correlating the echo patterns in a multiscale analysis, and filtering the noise sources, usually uncorrelated among different scales. Then, images with PC are differentiated by a binary classification method, evaluating Support Vector Machines and Adaboost models. Additionally, the proposed method is assessed using a public EUS database constructed and released in this work, with 55 cases. Finally, the proposed method was compared with typical Deep Learning approaches, reaching an accuracy of 92.1\% and 90.0\%, respectively. In addition, the method herein proposed is also stable in experiments with added noise, while the nets fail to maintain a similar performance.