Abstract: Geometric and mechanical properties of biological soft tissue are important indicators of many diseases. The properties of biological soft tissue acquired by high-precision inversion, such as tissue thickness, elastic modulus, density and so on, can provide significant medical reference to the diagnosis, treatment and recovery of diseases. The parameter information in ultrasonic detection can efficiently and accurately solve the parameter inversion problem. By combining dispersion curves of ultrasonic guided waves with machine learning, a back propagation neural network inversion method based on the Lamb wave dispersion curves of biological soft tissue was established in this paper. The effects of the dispersion mode, waveband, training data number, wavenumber sampling number and data noise on the inversion accuracy were analyzed. The results show that the inversion method combined dispersion curves and machine learning can accurately derive the single parameter properties of biological soft tissue. Moreover, this method has good universality and can be applied to the inversion of various characteristic parameters. The data of dispersion curves are obtained by solving the dispersion equation of biological soft tissue. The determination coefficientR² is introduced to evaluate the inversion results. Comparison among the inversion results using different modes and wavebands of dispersion curves proves that the dispersion mode and waveband of dispersion curves can be effectively and properly selected according to the sensitivity index obtained by the modified Morris method. Adopting mixed dispersion modes can further improve the inversion accuracy. When training data number and wavenumber sampling number reach certain thresholds, the inversion accuracy is basically unchanged. According to thresholds, both excellent inversion accuracy and inversion efficiency can be guaranteed. Robustness test was conducted by adding random noise to the sample data. The proposed inversion method in this paper shows considerable robustness when the training and testing samples have similar parameter value ranges and noise levels.