In subtraction microscopy, the negative sidelobes are inevitably generated by the difference between the envelopes of Gaussian and doughnut point spread functions (PSFs), resulting in undesired information loss. Therefore, the trade-off between high resolution and information loss hinders further improvement in the performance of subtraction microscopy. Moreover, the postprocessing subtraction algorithms derived from PSF algebra tend to cause artifacts in dense samples. Herein, we propose an adaptive algorithm for assignment of the subtractive coefficient based on deep learning, termed Deep-IWS, to enhance the performance of subtraction microscopy. Both simulation and experiment reveal that Deep-IWS increases the resolution 1.8 times better than confocal microscopy, and significantly outperforms the previous subtraction microscopy. Furthermore, the reconstructed images also have fewer artifacts with a higher signal-to-noise ratio (SNR), demonstrating the validity and superiority of our method.