Improving Medical Imaging Model Calibration through Probabilistic Embedding

Neural network model calibration is crucial in medical imaging, where accurate probabilistic predictions are essential for informed decision-making. Existing calibration techniques often introduce additional complexity and may not fully capture the inherent uncertainty associated with the tasks. To address these challenges, we propose a novel approach based on probabilistic embedding that models uncertainty through a Gaussian distribution. By embedding the model’s predictions into a probabilistic space, the proposed method enables effective uncertainty quantification. We demonstrate the effectiveness of our approach on multiple medical imaging tasks. The experimental result shows our method outperforms existing techniques in terms of both calibration and accuracy.