Due to the substantial feature extraction and end-to-end learning capability, deep learning has been widely used in intelligent medical image detection. However, amount of parameters in these models relies on the number of labeled training data, which influences the performance. Due to this reason, we propose a novel unsupervised medical image detection model named Multi-Scale Normalizing Flow (MS-NF). First, a fusion backbone network is applied to extract the multi-scale feature maps, which capture the different scale features of the anomalies. Second, normalizing flow transfers the abnormal distribution into the normal distribution hidden in the latent space, which is used for anomaly detection. To further improve the detection performance, channel and spatial convolutional attention mechanisms are integrated to make the model focus on the anomalous region by a shared network. Experimental results obtained on brain tumor MRI and ISIC2018 datasets show that MS-NF improves the pixel-level AUC index by 9% compared to existing medical image detection models, also performing well on small-scale data with efficient training and inference.