It is of great significance in clinical diagnosis and treatment evaluation to accurately identify the lesion tissue and further extract its characteristics and depth location. In this study, we construct tissue phantoms for three lesion types: fibrosis (FT), organelle proliferation (OPT), and pigmentation (PT). These phantoms allow for the quantitative regulation of mimicked disease depth. The experimental results show that the parameter Kc, combined with MMT parameters, can effectively distinguish the presence of lesions and their abnormal types. Further, the study extracts depth-sensitive polarization feature parameters (DSPFPs) for specific lesion types. Through experiments of tissue phantoms with various depth settings, the established machine learning regression models based on DSPFPs demonstrate their depth retrieval capabilities.