An improved low-rank and sparse matrix decomposition model was proposed to preserve edge, detail features and remove noise in ultrasound images. Firstly, logarithmic transformation was carried out to transform the multiplicative speckle noise into additive noise. Then the L1-norm and improved low-rank regularizer term were introduced to iteratively recover the denoised ultrasound image with the minimisation of the fidelity and regularization term as the objective function. Finally, the exponential transformation was used to restore the resulting graph from the logarithmic domain. The model was compared with some classical denoising algorithms using tumour ultrasound images. The results show that the model has good applicability and real time for the estimation of ultrasonic image denoising of gastrointestinal submucosal tumours.