Data scarcity is a common and challenging issue when working with Artificial Intelligence solutions, especially those including Deep Learning (DL) models for tasks such as image classification. This is particularly relevant in healthcare scenarios, in which data collection requires a long-lasting process, involving specific control protocols. The performance of DL models is usually quantified by different classification metrics, which may provide biased results, due to the lack of sufficient data. In this paper, an innovative approach is proposed to evaluate the performance of DL models when labeled data is scarce. This approach, which aims to detect the poor performance provided by DL models, in spite of traditional assessing metrics indicating otherwise, is based on information theoretic concepts and motivated by the Information Bottleneck framework. This methodology has been evaluated by implementing several experimental configurations to classify samples from a plantar thermogram dataset, focused on early stage detection of diabetic foot ulcers, as a case study. The proposed network architectures exhibited high results in terms of classification metrics. However, as our approach shows, only two of those models are indeed consistent to generalize the data properly. In conclusion, a new methodology was introduced and tested to identify promising DL models for image classification over small datasets without relying exclusively on the widely employed classification metrics.

Diabetes mellitus presents a high prevalence around the world. A common and long-term derived complication is diabetic foot ulcers (DFUs), which have a global prevalence of roughly 6.3%, and a lifetime incidence of up to 34%. Infrared thermograms, covering the entire plantar aspect of both feet, can be employed to monitor the risk of developing a foot ulcer, because diabetic patients exhibit an abnormal pattern that may indicate a foot disorder. In this study, the publicly available INAOE dataset composed of thermogram images of healthy and diabetic subjects was employed to extract relevant features aiming to establish a set of state-of-the-art features that efficiently classify DFU. This database was extended and balanced by fusing it with private local thermograms from healthy volunteers and generating synthetic data via synthetic minority oversampling technique (SMOTE). State-of-the-art features were extracted using two classical approaches, LASSO and random forest, as well as two variational deep learning (DL)-based ones: concrete and variational dropout. Then, the most relevant features were detected and ranked. Subsequently, the extracted features were employed to classify subjects at risk of developing an ulcer using as reference a support vector machine (SVM) classifier with a fixed hyperparameter configuration to evaluate the robustness of the selected features. The new set of features extracted considerably differed from those currently considered state-of-the-art but provided a fair performance. Among the implemented extraction approaches, the variational DL ones, particularly the concrete dropout, performed the best, reporting an F1 score of 90% using the aforementioned SVM classifier. In comparison with features previously considered as the state-of-the-art, approximately 15% better performance was achieved for classification.

<p>The Contrastive Learning for blind Hyperspectral Unmixing (CLHU) is a self-supervised deep learning approach for blind hyperspectral unmixing. Unlike existing deep learning methods that rely on reconstruction capabilities, CLHU leverages the input-endmembers relationship for abundance estimation. The endmembers are dynamically updated during training, controlled by two regularization factors. Extensive experiments demonstrate CLHU's effectiveness, achieving state-of-the-art performance in hyperspectral unmixing. This novel approach offers a promising perspective for the field and holds potential for further enhancements in hyperspectral unmixing tasks.</p>

<p>The Contrastive Learning for blind Hyperspectral Unmixing (CLHU) is a self-supervised deep learning approach for blind hyperspectral unmixing. Unlike existing deep learning methods that rely on reconstruction capabilities, CLHU leverages the input-endmembers relationship for abundance estimation. The endmembers are dynamically updated during training, controlled by two regularization factors. Extensive experiments demonstrate CLHU's effectiveness, achieving state-of-the-art performance in hyperspectral unmixing. This novel approach offers a promising perspective for the field and holds potential for further enhancements in hyperspectral unmixing tasks.</p>