In medical image analysis, a reliable model is required to detect inputs containing important anatomical information and make accurate decisions based on it. Motivated by this, we introduce the concept of “out-of-clinical-distribution” (OCD) detection, where in-clinical-distribution data (ICD) is defined as images containing a “clinically interesting region” that is essential to clinical decision-making. We propose an OCD detection framework based on a classification-based model, enhanced by a novel softmax-conditioned variational autoencoder regulariser. In this framework, softmax scores are incorporated into the latent space with a mixture of learnable class-conditioned Gaussian distributions as prior. By embedding class information in feature reconstruction, this approach enforces feature compactness within ICD classes and enhances the separability between ICD and OCD features. The effectiveness of the proposed OCD detection method is demonstrated in the task of selecting anatomical views from real-time fetal ultrasound (US) videos, where it significantly outperforms both state-of-the-art classification-based and generative-based methods.