The role of medical imaging in enhancing patient outcomes through early detection is becoming increasingly important. This research work focuses on optimizing liver cancer detection from MRI scans using advanced preprocessing techniques and deep learning models. It indicates the integration of the traditional preprocessing techniques with feature extraction based on convolutional neural networks (CNNs) and noise reduction based on convolutional denoising autoencoders (CDAEs). Four deep learning models, namely ResNet-50, DenseNet, VGG-19 and Inception V3, are trained and tested using the two sets of data, one of which had been applied with the traditional pre-processing and the other set with the hybrid pre-processing. This outcome shows that the hybrid preprocessing method improved the accuracy levels of the model as ResNet-50 recorded the highest accuracy at 98.9%, followed closely by DenseNet with 96.7%, VGG-19 with 93.4%, and Inception V3 with 91.2% accuracy levels when trained using the hybrid method. Nevertheless, if the traditional preprocessing methods were applied, performance levels for all the models were lower; this is for ResNet-50 at 95.6%, DenseNet 90.23%, VGG-19 88.76%, and Inception V3 95.4%. These results support that the hybrid preprocessing pipeline is indeed effectively enhancing the performance of deep learning models in achieving good accuracy levels in terms of the detection of liver cancer. The methodology proposed herein holds great promise for real-time medical imaging applications and provides a reliable, efficient, and early tool for clinical decision-making. This research opens up avenues for further extensions in the region of performance amelioration in medical image analysis using more enhanced AI techniques.