| Abstrakt: |
Lymphoma is a type of malignant tumor that develops from lymphoid hematopoietic tissues. The precise diagnosis of lymphomas is one of the challenging tasks because of the similarity within the morphological features across lymphoma classes. Therefore, an effective classification of lymphoma is required to be developed for taking a timely action for the patients. This paper aims to propose the Gradient Propagation based DenseNet121 namely GPDN for an effective multiclass lymphoma classification. The dense connections of GPDN is used to mitigate the vanishing gradient issue by confirming an effective gradient propagation over the network for enhancing stability and avoiding overfitting issue. An effective reutilization of features from previous layers, the GPDN obtains the intricate for enhanced differentiation of lymphoma. Different pre-processing approaches such as gaussian filter, min-max normalization and data augmentation are developed for denoising, rescaling and augmenting the input images. Here, multiscale features from the image are obtained by using ResNet50-based feature extraction. This research uses a dataset from Kaggle named Multi cancer dataset. The performance of the proposed ResNet50-GPDN is analyzed using accuracy, precision, recall and F-measure. Simulation results show that the GPDN achieves higher accuracy of 99.90% than the VGG16, InceptionV3, MobileNet and AlexNet. Moreover, existing researches HPC and FFNN-ResNet50, are used for the comparison of the ResNet50-GPDN. The accuracy of ResNet50-GPDN is 99.90% which is clearly higher than those of HPC and FFNN-ResNet50. |
Full text from SpringerLink