Objective <p>Recently, deep learning models in medical imaging have undergone tremendous advancements. However, these deterministic discriminative models often tend to overfit and produce overconfident predictions. To explore a more efficient strategy, we propose a diffusion-conditioned representation learning that leverages the internal dynamics of a diffusion model to extract noise-level awareness and diverse semantic representations in disease classification and detecting distributional shifts. The proposed method inherently encodes aleatoric uncertainty by sampling features from multiple denoising time steps, which is then used to regularize prediction during training via entropy regularized loss.</p> Result <p>We conducted extensive experimentation over five diverse clinical datasets, and our framework demonstrates competitive performance with reduced prediction entropy and improved robustness to input degradation. The negative entropy–accuracy correlation, expected calibration error, and dynamic uncertainty demonstrate the effectiveness of the proposed framework in classification tasks. Our framework achieved competitive performance compared with traditional convolutional neural networks and vision transformers with an AUC of 83.81%, 97.45%, 86.15% and 96.75% on datasets 1, 2, 3 and 4, respectively. The reduced ECE values on the datasets demonstrate the calibrated performance of the proposed method.</p>