Weakly- and Semi-supervised Graph CNN for Identifying Basal Cell Carcinoma on Pathological Images

• Published in: Graph Learning in Medical Imaging Vol. 11849; pp. 112 - 119
• Main Authors: Wu, Junyan, Zhong, Jia-Xing, Chen, Eric Z., Zhang, Jingwei, Ye, Jay J., Yu, Limin
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2019; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Basal Cell Carcinoma; Graph-CNN; Pathology
• ISBN: 303035816X; 9783030358167
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-030-35817-4_14

Deep learning has been used to identify Basal Cell Carcinoma (BCC) from pathology images. The traditional patch-based strategy has the problem of integrating patch level information into the whole image level prediction. Also, it is often difficult to obtain sufficient high-quality patch labels such as pixel-wise segmentation masks. Benefiting from the recent development of Graph-CNN (GCN), we propose a new weakly- and semi-supervised GCN architecture to model patch-patch relation and provide patch-aware interpretability. Integrating prior knowledge and structure information, without relying on pixel-wise segmentation labels, our whole image level prediction achieves state-of-art performance with mAP 0.9556 and AUC 0.9502. Further visualization demonstrates that our model is implicitly consistent with the pixel-wise segmentation labels, which indicates our model can identify the region of interests without relying on the pixel-wise labels.