DEEP LEARNING BASED LYMPHOMA DETECTION ON MEDICAL IMAGES BY USING ENHANCED RESNET AND VGG NETWORKS

Abstract

Lymphoma is a key hematologic malignancy where precise and fast diagnostic interventions are elicited to guide effective treatment interventions. The traditional histopathological analysis is lengthy and prone to interobserver variability, and thus hinders quick clinical decision making. This paper presents a framework of the automated identification of lymphoma using deep learning and adjusted VGG16 and ResNet50 networks. Transfer learning, global average pooling, custom fully connected layers and dropout regularization are used to augment the models in order to improve generalizability. The validation based on the experiment is carried out on a publicly available histopathological dataset, that is, Chronic Lymphocytic Leukaemia (CLL), Follicular Lymphoma (FL) and Mantle Cell Lymphoma (MCL). The four metrics used to evaluate performance are accuracy, precision, recall, F1 -score and Kappa as defined by Cohen. The modified VGG16 had the highest accuracy of 97.65 0 -1, and the modified ResNet50 had the highest accuracy of 96.40 0 -1, which proves that there is good performance in the classification.  Statistical validation confirms that the improvements are significant (p < 0.05). These were results highlighting the effectiveness of the internalization and optimization of architectural design with dependable lymphoma differentiation. The suggested structure has great translational capability to incorporate it into the computer-aided diagnosis systems.