At present, image segmentation is widely employed to identify areas of an object. Although most input images can be fit into the GPU memory, in some applications, such as lung histopathology images, each image has more than 1 gigapixels and it is impossible to fit the entire image into the GPU memory. In such a situation, we need to divide the image into sub-images and collectively segment each sub-region. In this work, we study behavior of the image segmentation in sub-images for wound segmentation. This work analyzes how changes in the training areas affect the accuracy of segmentation when results from sub-images are merged back to the original image size. In other words, we can tune the accuracy by sampling some regions more frequently to produce training data focusing on challenging regions. Our analysis showed that some regions of a target object could be harder to segment and providing sub-images from these regions tended to improve overall accuracy. We also propose to use accuracy heat map to visualize impacts of sampling subregions on the accuracy. The results of adding appropriate data to focus on challenging areas in which the machine has low accuracy problems improves performance in terms of intersection over union (IoU) as follows. When a machine was trained with 10 images, the IoU increased from 52.23% to 53.79%. When trained with 20 images and 115 images, IoU increased from 61.46% to 64.59%, and 77.97% to 78.16%. The proposed technique can be generalized and applied to other segmentation problems.