Compton imaging is a promising method of sub MeV to a few MeV gamma-rays and expected to use in various application field, such as medical imaging, environmental monitoring and astrophysics. Several types of Compton camera has beed developed using different materials. One of the drawbacks in conventional Compton imaging is relatively low signal to background ratio caused by its projected Compton cones. Recoil electron tracking is one straight-forward way to improve the signal-to-background ratio, however, it is only realized in gaseous detectors. The realization of electron tracking in solid detectors is under investigation because of its short track in scatter materials. We demonstrated the capability of electron tracking in silicon-on-insulator (SOI) pixel detector with 30 μm pixels size. The extraction of ejected direction of recoil electrons in Compton scattering is an important problem. In this work, we investigated the use of deep learning for estimating the angle in plane and depth using Geant 4 Monte Carlo simulation. The effect of pixel size to the estimation accuracy and SPD is characterized for the application of actual silicon-on-insulator based SOI detectors. The imaging capability is also characterized using predicted recoil electron direction.