Liver cancer is a major cause of cancer-related mortality worldwide, and the accurate segmentation of small tumors in CT scans remains a critical challenge due to their morphological variability and indistinct boundaries. To address this, we propose MSCAM-Net, a Multi-scale Cross-Attention Modulation Network specifically designed to enhance small liver tumor segmentation. The network integrates a Spatial-Channel Attention Atrous Spatial Pyramid Pooling (SCAASPP) module to extract rich multi-scale contextual features and a Spatial-Enhanced Bi-Directional Cross Modulation Decoder (SBCMD) to adaptively fuse semantic and spatial information through dual-path interaction. This design enables MSCAM-Net to preserve global coherence while capturing fine-grained details crucial for delineating small, low-contrast lesions. Evaluated on the LiTS dataset, MSCAM-Net consistently outperforms state-of-the-art models in segmentation metrics such as Dice, VOE, RVD, and ASD. Visual results further validate its effectiveness, underscoring its potential for clinical application in early liver cancer detection and personalized treatment planning.