Abstract: In this study, we present a lightweight, nondestructive testing model for diagnosing common diseases in crucian carp as an example. This model, based on an improved YOLOv5s architecture, aims to achieve simultaneous, nondestructive and rapid detection of multiple diseases in crucian carp, addressing the practical needs of the industry. The proposed model is designed to be lightweight and is enhanced by replacing the YOLOv5s backbone network with ShuffleNetV2. Additionally, it incorporates a convolutional block attention module (CBAM) to enhance model accuracy. Furthermore, the model is combined with the cavity space convolutional pooling pyramid (ASPP) to improve model robustness. Through testing on the homemade crucian carp disease dataset, it can be seen that the disease detection accuracy of the model proposed can reach 92.0%, and the volume of the model is only 14400 kb, which is better than the current mainstream model (the highest accuracy is 83.6%, and the smallest volume is 15750 kb), which provides technical support for the non-destructive and rapid detection of fish diseases in aquaculture.