Optical lithography is a fundamental process to fabricate integrated circuits, which are the basic fabric of the information age. Due to image distortions inherent in optical lithography, inverse lithography techniques (ILT) are extensively used by the semiconductor industry to improve lithography image resolution and fidelity in semiconductor fabrication. As the density of integrated circuits increases, computational complexity has become a central challenge in ILT methods. This paper develops a new and powerful framework of a model-driven convolution neural network (MCNN) to obtain the approximate guess of the ILT solutions, which can be used as the input of the following ILT optimization with much fewer iterations as compared with conventional ILT algorithms. The combined approach to use the proposed MCNN together with the gradient-based method can improve the speed of ILT optimization algorithms up to an order of magnitude and further improve the imaging performance of coherent optical lithography systems. This paper, to the best of our knowledge, is the first to exploit a state-of-the-art MCNN to solve the ILT problem and provide considerable performance advantages. The imaging model of optical lithography is utilized to establish a neural network architecture and an unsupervised training strategy. The neural network architecture and the initial network parameters are derived by unfolding and truncating the model-based iterative ILT optimization procedure. Then, a model-based decoder is proposed to enable the unsupervised training of the neural network, which averts the time-consuming labelling process in training data. This work opens a new window for MCNN techniques to effectively improve the computational efficiency and imaging performance of conventional ILT algorithms. Some impressive good simulation results are provided to verify the superiority of the proposed MCNN approach.