Detecting voice disorders attracts increasing research interests. This paper reports the performance results of two-dimensional (2D) speech representations when used as inputs to a convolutional neural network (CNN)-based system. The 2D representations investigated in this paper are spectrograms, and three human auditory perception-scaled spectrograms namely, octave-scaled, mel-scaled, and gammatone-scaled spectrograms. To improve the performance of these representations, an innovative voice enhancement approach is integrated within a deep learning process. The new enhancement-based approach improved the performance of gammatone-scaled spectrograms and octave-scaled spectrograms by achieving an unweighted average recall (UAR) of 73.29% and 74.04%, respectively, using the Saarbruecken voice database (SVD).