| Abstrakt: |
Convolutional Neural Network (CNN) methods are employed in medical image analysis for computer-aided diagnosis. CNNs are state-of-the-art in many image-related applications, however, they lose spatial information between picture occurrences, need large training sets, and have poor input transformations. CNN's limitations can be overcome through CapsNet via encode and decode operations. However, CapsNets are emerging for medical image categorization and need complexity reduction optimizations. We present Time Efficient-CapsNet (TE-CapsNet), a unique model, to classify diseases accurately with low computing complexity by utilizing medical images. Time-efficient feature estimation in the encoder is the main contribution to TE-CapsNet. It is achieved by modifying the conventional CapsNet layers using different pre-trained CNN models and activation functions. Before applying the TE-CapsNet to input medical images, we perform the vital steps of medical image pre-processing and segmentation to enhance classification accuracy. To minimize the time complexity, we design the minimum CNN layers and CapsNet layers to process the maximum-sized medical image in the TE-CapsNet model. We propose the TE-CapsNet model for a problem of medical disease classification from input medical images with higher accuracy and lower computational requirements. The TE-CapsNet framework is designed and evaluated using two publicly available medical image datasets of small size. According to achieved results, the TE-CapsNet model's accuracy, precision, recall, F1-score performance, and specificity rates have all improved by approximately 3.5% compared to CNN models and state-of-the-art. The computational complexity is reduced by 11% using the proposed model. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink