Generalized Zero-Shot Learning (GZSL) addresses the challenge of classifying unseen classes in the presence of seen classes by leveraging semantic attributes to bridge the gap for unseen classes. However, in image based disease classification, such as glioma sub-typing, distinguishing between classes using image semantic attributes can be challenging. To address this challenge, we introduce a novel GZSL method that eliminates the dependency on semantic information. Specifically, we propose that the primary of most classification in clinic is risk stratification, and classes are inherently ordered rather than purely categorical. Based on this insight, we present an inter-class feature augmentation (IFA) module, where distributions of different classes are ordered by their risk levels in a learned feature space using pre-defined joint conditional Gaussian distribution model. This ordering enables the generation of unseen class features through feature mixing of adjacent seen classes, effectively transforming the zero-shot learning problem into a supervised learning task. Our method eliminates the need for explicit semantic information, avoiding the cross-modal alignment between visual and semantic features. Moreover, the IFA module for GZSL requires no structural modifications to the existing classification models. In the experiment, both in-house and public datasets are used to evaluate our method across different tasks, including glioma subtyping, Alzheimer’s disease (AD) classification and diabetic retinopathy classification. Experimental results demonstrate that our method outperforms the state-of-the-art GZSL methods with statistical significance.

Autism spectrum disorder (ASD) is a complex neurodevelopmental disorder severely impacting social abilities. Previous research has demonstrated impairments in functional brain connections among individuals with ASD. This paper proposed a novel diagnostic strategy for ASD using spatial-temporal deep neural networks (DNN) integrating the self-attention mechanism. Specifically, we employed sliding window method in the temporal dimension to construct dynamic functional connectivity (dFC) data and applied Kendall’s rank correlation coefficient to extract features. Subsequently, deep neural networks based on multi-head self-attention mechanism were designed to extract more discriminative spatial and temporal high-level abstract features. The learned spatial-temporal features from multiple layers of multi-head self-attention were concatenated and subsequently incorporated with feed-forward neural networks. We designed a novel disagreement regularization term combined with cross-entropy loss for classification. We conducted systematic experiments on the large-scale ASD dataset. Both 10-fold and inter-site cross-validation results outperform existing classical studies in ASD classification, suggesting the effectiveness of the proposed model. Moreover, we identified the discriminative functional connections associated with ASD classification.