Prestack amplitude variation with offset (AVO) inversion using one-dimensional convolutional neural networks often lacks lateral continuity. While two-dimensional methods improve this, they are limited to unidirectional spatial correlations from well to non-well locations. To overcome these limitations, we propose a semi-supervised learning approach with bidirectional spatial feature constraints (BSFC-SSL). Our method introduces a label-annihilation operator and a dedicated spatial feature network to establish bidirectional information flow between well and non-well locations, thereby capturing more complex spatial patterns in seismic data. Integrated with semi-supervised learning and low-frequency constraints, the BSFC-SSL framework enhances both stability and generalization. Experiments on synthetic and field data demonstrate that our method achieves superior lateral continuity and inversion accuracy compared to conventional one- and two-dimensional deep learning techniques.