An automatic visual inspection method based on supervised machine learning for rapid on-site evaluation in EUS-FNA

In this paper, an automatic visual inspection method based on supervised machine learning is proposed to assist rapid on-site evaluation (ROSE) for endoscopic ultrasound-guided fine needle aspiration (EUS-FNA) biopsy. The aim of this method is to learn relations between content of cellular tissue including tumor cells and aspect of specimen image removed by the needle aspiration. For this purpose, a stationary Gaussian mixture model (GMM) is applied to classify the local statistics of the specimen images, because stationary GMM is known to be effective to estimate universal model. In this paper, some specimen images with their definitive diagnosis information are used as training images in GMM learning. The training images are also used in the supervised learning with their diagnosis information as teacher data, i.e. the rank of tumor cells content. Thus, the learning of statistical relation between the local image aspect and its rank of tumor cells content may be linked by the class index of GMM, using the training images. A simulation result shows that the proposed method is effective to assist on-site visual inspection of cellular tissue in ROSE for EUS-FNA, indicating highly probable area including tumor cells.