Breast tumors threaten women’s health. Multispectral transmission imaging shows potential in early breast cancer screening, but images have a low signal-to-noise ratio due to tissue absorption and scattering. Frame accumulation improves quality but leads to data redundancy. Existing terraced compression methods (TCM) address this but rely on subjective, inefficient threshold selection. This paper presents an improved TCM that requires only the input of the desired heterogeneity size, with optimal thresholds adaptively determined by analyzing grayscale values and their distribution. By applying compression to each wavelength image and using the GMM algorithm for multidimensional image clustering, results show the Dice coefficient improved by 12.49 %, and the recognition rate of deeper heterogeneity increased by 29.09 % compared to the existing TCM. This validates the TCM’s effectiveness and threshold selection accuracy. Additionally, the method distinguishes heterogeneous bodies of different depths and effectively recognizes various kinds, providing a reliable reference for early breast cancer screening.