Fusion of Hand-crafted and Deep Features for Automatic Diabetic Foot Ulcer Classification

Abstract

This paper proposes to combine both the texture and deep features to build a robust diabetic foot ulcer recognition system since both features represent valuable information about the disease. The proposed system consists of three stages: feature extraction, feature fusion, and DFU classification. The feature extraction is performed by extracting the handcrafted and deep features. The feature fusion is performed by concatenating both feature vectors into a single vector. The DFU classification is performed by training a random forest classifier on the fusion vectors and the resulting classifier is used then for classification. Experimental results showed that the proposed approach provides satisfactory performance in DFU, ischaemia, and infection classification.

Towards Home-Based Diabetic Foot Ulcer Monitoring: A Systematic Review

“…Though the classification used achieved the poorest results (F1 score: 0.942 (wound), 0.99 (ischemia), and 0.744 (infection)), the smallest dataset was used for model training. Similar results were achieved (F1 score: 0.89 (wound), 0.93 (ischemia), and 0.76 (infection)) [43] by fusing features extracted by using deep (GoogLeNet [73]) and machine learning algorithms (Gabor [82] and HOG [83]) and applying a random forest classifier. It was demonstrated that fused features outperformed deep features alone.…”

“…One of the public datasets [ 40 ] contained augmented data that were preprocessed before augmentation. As a result, the articles that used this dataset did not used any preprocessing or augmentation techniques [ 17 , 42 , 43 , 44 , 45 ].…”

“…Though all of the reviewed articles described their methodologies, for a few of the articles, it was difficult to objectively evaluate the backbones’ performance. Some studies presented binary classification results for multiple different classes [ 40 , 41 , 43 , 44 , 45 , 46 ], while others used unpopular performance metrics [ 22 , 23 , 58 ] and were excluded from this analysis. It could be observed that a performance metric of 0.9 or higher was achieved by using custom architectures, Inception, ResNet152, and other pretrained networks.…”