Automatic Detection of Common Gastroenterological Diseases Using a Small Dataset: A Two-Phase Image Processing Method

Manually analyzing images from LASER Confocal Endomicroscopy (CLE) images is a common approach for diagnosing gastroenterological diseases that requires training and experience. The subjective aspect of this approach may lead to misdiagnoses, especially when the objective is to identify which of the diseases that can lead to esophageal cancer the patient possess: Gastric Metaplasia, Barret’s Esophagus, or Neoplastic Mucosa. Automating the diagnosis has been explored by some authors, but without regard for specific disease identification. This work proposes a classification method for images obtained from CLE exams that combine results of two classifiers based on features obtained from Gray Level Co-occurrence Matrices and Local Binary Patterns, two traditional texture description methods. The approach divides a CLE image into four regions of interest and then classifies each region with both classifiers. The final step combines the results of both classifiers using a consensus algorithm into a final prediction of the original image. The proposed method could classify the images by the disease they represent, not only by whether they contain or not a disease, which is the most common approach by similar methods. The method also achieved a result comparable to related work despite using a much smaller dataset for training.