In photoacoustic tomography, an ultrasonic transducer array is usually used to receive photoacoustic signals, which is expensive to manufacture, and the number of array elements has an important impact on the final imaging quality. To improve photoacoustic image quality reconstructed under sparse view conditoin, this study proposes a modified U-Net based on the replacement of the skip connection in a conventional U-Net with continuous convolutional layers, thereby increasing the matching degree of features transferred from the encoder to the decoder. Furthermore, the loss function based on the structural similarity index measure is used to train the network. Experimental results based on simulation and in vivo dataset show that compared with the conventional U-Net, the modified U-Net achieves more image details and the quality of the reconstructed image is significantly better.