Authors:
Jia Qu
1
;
Nobuyuki Hiruta
2
;
Kensuke Terai
2
;
Hirokazu Nosato
3
;
Masahiro Murakawa
1
;
3
and
Hidenori Sakanashi
1
;
3
Affiliations:
1
Department of Intelligent Interaction Technologies, University of Tsukuba, Tsukuba, Japan
;
2
Department of Surgical Pathology, Toho University Sakura Medical Center, Sakura, Japan
;
3
Artificial Intelligence Research Center, National Institute of Advanced Industrial Science and Technology (AIST), Japan
Keyword(s):
Pathology Image, Deep Learning, Transfer Learning, Color-Index Local Auto-Correlation (CILAC).
Abstract:
Deep learning using Convolutional Neural Networks (CNN) has been demonstrated unprecedentedly powerful for image classification. Subsequently, computer-aided diagnosis (CAD) for pathology image has been largely facilitated due to the deep learning related approaches. However, because of extremely high cost of pathologist's professional work, the lack of well annotated pathological image data to train deep neural networks is currently a big problem. Aiming at further improving the performance of deep neural networks and alleviating the lack of annotated pathology data, we propose a full-automatic knowledge transferring based stepwise fine-tuning scheme to make deep neural networks follow pathologist’s perception manner and understand pathology step by step. To realize this conception, we also introduce a new type of target correlation intermediate dataset which can be yielded by using fully automated processing. By extracting rough but stain-robust pathology-related information from u
nannotated pathology images with handcrafted features, and making use of these materials to intermediately train deep neural networks, deep neural networks are expected to acquire fundamental pathological knowledge in advance so that boosted in the final task. In experiments, we validate the new scheme on several well-known deep neural networks. Correspondingly, the results present solid evidence for the effectiveness and suggest feasibility for other tasks.
(More)