National Science and Technology Support Program [2012BAI14B03]
; National Natural Science Foundation of China [61375014, 61533015, 61333019]
; Foundation of Chinese Scholarship Council
Goal: Most state-of-the-art computer-aided endoscopic diagnosis methods require pixelwise labeled data to train various supervised machine learning models. However, it is a tedious and time-consuming work to collect sufficient precisely labeled image data. Fortunately, we can easily obtain huge endoscopic medical reports including the diagnostic text and images, which can be considered as weakly labeled data. Methods: In this paper, ourmotivation is to design a new computer-aided endoscopic diagnosis system without human specific labeling; in comparison with most state of the arts, ours only depends on the endoscopic images with weak labels mined from the diagnostic text. To achieve this, we first cast the endoscopic image folder and included images as bag and instances and represent each instance based on the global bag-of-words model. We then adopt a feature mapping scheme to represent each bag by mining the most suspicious lesion instance from each positive bag automatically. In order to achieve self-online updating from sequential new coming data, an online metric learning method is used to optimize the bag-level classification. Results: Our computer-aided endoscopic diagnosis system achieves an AUC of 0.93 on a new endoscopic image dataset captured from424 volunteers with more than 12k images. Conclusion: The system performance outperforms other state of the arts when we mine the most positive instances from positive bags and adopt the online phase to mine more information from the unseen bags. Significance: We present the first weakly labeled endoscopic image dataset for computer-aided endoscopic diagnosis and a novel system that is suitable for use in clinical settings.