Hamed Aghili - Department of Computer and Information technology (Robotic engineering),Payame Noor University (PNU),IRAN

Inspired by the success of deep learning architectures, especially deep convolutional neural networks (CNNs) in different machine learning and image classification tasks, in this work, these structures are applied for histopathological image classification. In particular, transfer learning of deep models to the medical image analysis domain is investigated. Transferring knowledge from other domains to that of histopathological images is motivated by the significantly lower number of histopthodological images for training as compared with other general images in addition to the computationally expensive training stage of deep networks. In order to investigate the possibility of transferring such knowledge, different deep nets, pre-trained on nonmedical image data are examined for classification purposes. All models evaluated are CNN structures which are trained with a wide variety of non-medical images. For the purpose of this study, we have examined eighteen state-of-the-art pretrained deep models and identified the best ones for classification of histopathological images. The experiments are conducted on a mammalian histopathological image database provided by Animal Diagnosis Lab (ADL) fromPennsylvania State University. ADL is a challenging dataset which consists of three bovine organs (kidney, lung, and spleen). The experiments revealed that deep pre-trained models can achieve great performance in classification of histopathological images. The best performing deep networks are then identified and compared with the state-of-art methods for classification of histopathological images, demonstrating the viability of transferring knowlwdge from non-medical domains to that of histopathological images with greate success. In particular, the pre-trained models have outperformed the state-of-the-art methods by a large margin.

Transfer Learning, Deep learning, Convolutional Neural Networks, Histopathological Images, Medical Image Diagnosis