Title: Deep Subspace analysing for Semi-Supervised multi-label classification of Diabetic Foot Ulcer

Abstract: Diabetes is a global raising pandemic. Diabetes patients are at risk of developing foot ulcer that usually leads to limb amputation. In order to develop a self monitoring mobile application, in this work, we propose a novel deep subspace analysis pipeline for semi-supervised diabetic foot ulcer mulit-label classification. To avoid any chance of over-fitting, unlike recent state of the art deep semi-supervised methods, the proposed pipeline dose not include any data augmentation. Whereas, after extracting deep features, in order to make the representation shift invariant, we employ variety of data augmentation methods on each image and generate an image-sets, which is then mapped into a linear subspace. Moreover, the proposed pipeline reduces the cost of retraining when more new unlabelled data become available. Thus, the first stage of the pipeline employs the concept of transfer learning for feature extraction purpose through modifying and retraining a deep convolutional network architect known as Xception. Then, the output of a mid-layer is extracted to generate an image set representer of any given image with help of data augmentation methods. At this stage, each image is transferred to a linear subspace which is a point on a Grassmann Manifold topological space. Hence, to perform analyse them, the geometry of such manifold must be considered. As such, each labelled image is represented as a vector of distances to number of unlabelled images using geodesic distance on Grassmann manifold. Finally, Random Forest is trained for multi-label classification of diabetic foot ulcer images. The method is then evaluated on the blind test set provided by DFU2021 competition, and the result considerable improvement compared to using classical transfer learning with data augmentation.