Deep learning approach for blur detection of digital breast tomosynthesis images

Image quality is critical in domains such as computer vision, image processing, and pattern recognition. One of the areas of image processing where image quality is critical is image restoration. In the field of medical imaging, blur detection is used in the pre-processing stage of medical image restoration. It was noted that blurring has the potential to obscure small cancers and microcalcifications. As a result, some abnormalities were undiscovered until they have grown significantly. The quality of an image can be determined whether it is blurry using various blur detection algorithms. This paper presents a comparative study of various pre-trained convolutional neural networks (CNNs) models as feature extraction for blur detection. The CNNs models are ResNet18, ResNet50, AlexNet, VGG16 and InceptionV3. These CNNs were then connected to a classifier known as support vector machine (SVM) to classify DBT images into blurry or not blurry. To evaluate the performance of the pre-trained CNN as features extractor, statistical performance measures namely the accuracy, receiver- operating characteristics (ROC), area under the curve (AUC), and execution time were employed. According to the evaluation results, InceptionV3 has the best accuracy rate at 0.9961 with AUC of 0.9961. Most of the output of Pre-trained CNN with SVM lies closest to the ideal ROC curve near the top left corner. AlexNet has the shortest processing time of any of the CNNs model. The findings of this study might be used as reference before performing image restoration.