Magnetic induction tomography is a new contactless electromagnetic imaging technique with uses in non-invasive testing, geophysics, and medical imaging. MIT suffers from two problems, the forward and inverse problem, both of which impact MITs resolution. A mixture of traditional and intelligent algorithms has been proposed to solve these problems. This paper proposes an intelligent, data-driven, deep learning model to solve both problems by computing the non-linear relationship between the voltage differences across the object and the top-down image of the object. An investigation is conducted into six classification and seven generative models to determine the optimum model. All these models are trained on two datasets of measured samples. One contains aluminium samples with vary shape complexity and the other contains copper samples with decreasing size. A variation of the well-known U-Net model was deemed the most optimal at generating binary images of objects with vary shapes and sizes with mean absolute error of 5.6601e-5. However, the model was unable to generate images of objects with a size smaller than 1600 mm2 or a side length of 40 mm. The model’s robustness was tested on three special cases containing samples of unknown shapes, unknown number of samples on the MIT sensor, and unknown object rotation. Though the model underperformed with these special cases, the investigation presents room for future work . The research proved that deep learning models can be used to generate images of sampled models in MIT and therefore provide a solution to the forward and inverse problems.