Semi-supervised medical image classification via feature-level multi-scale consistency and adversarial training

In recent years, semi-supervised learning (SSL) has attracted increasing attention in medical image analysis, showing great potential in scenarios with limited annotations. However, existing consistency regularization methods suffer from several limitations: overly uniform constraints at the output layer, lack of interaction within adversarial strategies, and reliance on external sample pools for sample estimation, which together lead to insufficient use of feature-level information and unstable training. To address these challenges, this paper proposes a novel semi-supervised framework, termed Feature-level multi-scale Consistency and Adversarial Training (FCAT). A multi-scale feature-level consistency mechanism is introduced to capture hierarchical structural representations through cross-level feature fusion, enabling robust feature alignment without relying on external sample pools. To overcome the limitation of unidirectional adversarial training, a bidirectional feature perturbation strategy is designed under a teacher-student collaboration scheme, where both models generate perturbations from their own gradients and enforce mutual consistency. In addition, an intrinsic evaluation mechanism based on entropy and complementary confidence is developed to rank unlabeled samples according to their information content, guiding the training process toward informative hard samples while reducing overfitting to trivial ones. Experiments on the balanced Pneumonia Chest X-ray and NCT-CRC-HE histopathology datasets, as well as the imbalanced ISIC 2019 dermoscopic skin lesion dataset, demonstrate that our FCAT achieves competitive performance and strong generalization across diverse imaging modalities and data distributions.