Glaucoma is a disease in which the optic nerve is chronically damaged by the elevation of the intra-ocular pressure, resulting in visual field defect. Therefore, it is important to monitor and treat suspected patients before they are confirmed with glaucoma. In this paper, we propose a 2-stage ranking-CNN that classifies fundus images as normal, suspicious, and glaucoma. Furthermore, we propose a method of using the class activation map as a mask filter and combining it with the original fundus image as an intermediate input. Our results have improved the average accuracy by about 10% over the existing 3-class CNN and ranking-CNN, and especially improved the sensitivity of suspicious class by more than 20% over 3-class CNN. In addition, the extracted ROI was also found to overlap with the diagnostic criteria of the physician. The method we propose is expected to be efficiently applied to any medical data where there is a suspicious condition between normal and disease.