In ophthalmology, retinal optical coherence tomography (OCT) images with noticeable structural features help identify human eyes as healthy or diseased. The recently hot artificial intelligence (AI) realized this recognition process automatically. However, speckle noise in the original retinal OCT image reduces the accuracy of disease classification. This study presents a time-saving approach based on deep learning to improve classification accuracy by removing the noise from the original dataset. Firstly, four pre-trained convolutional neural networks (CNNs) from the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) were trained to classify the original images into two categories: The noise reduction required (NRR) and the noise-free (NF) images. Among the CNNs, VGG19_BN performed best with 98% accuracy and 99% recall. Then, we used the block-matching and 3D filtering (BM3D) algorithm to denoise the NRR images. Those noise-removed NRR and the NF images form the processed dataset. The quality of images in the dataset is prominently ameliorated after denoising, which is valid to improve the models’ performance. The original and processed datasets were tested on the four pre-trained CNNs to evaluate the effectiveness of our proposed approach. We have compared the CNNs, and the results show the performance of the CNNs trained with the processed dataset is improved by an average of 2.04%, 5.19%, and 5.10% under overall accuracy (OA), Macro F1-score, and Micro F1-score, respectively. Especially for DenseNet161, the OA is improved to 98.14%. Our proposed method demonstrates its effectiveness in improving classification accuracy and opens a new solution to reduce denoising time-consuming for large datasets.