Capture the Devil in the Details via Partition-then-Ensemble on Higher Resolution Images

Abstract

For the Diabetic Foot Ulcer Challenge 2022 (DFUC2022) hosted by MICCAI 2022, we built a machine learning model based on the architecture of TransFuse [20] to accomplish the segmentation task. The TransFuse model combines Transformers and convolutional neural networks (CNNs), taking advantage of both local and global features. In this paper, we propose a modification to the data flow in encoder necks for decoding features in the higher resolution level, and in fusion modules for more efficient attention. Furthermore, to minimize the information loss as a result of resizing, we propose new techniques in both training and testing algorithms. Firstly, a region proposal network (RPN) is introduced from object detection methods and is used at the image pre-processing phase. It crops fixed size images from origin images, so that the high resolution input can be fed into TransFuse. We also applied test-time augmentation following a similar concept to RPN. We crop fixed size images at each corner and use edge pooling to ensemble them properly.