Comparative Study and Detection of Lymphoma through Medical Imaging and modified ResNet and VGG Models

Abstract

Diagnosis and treatment of lymphoma, an aggressive neoplasm of the lymphatic system, becomes extremely difficult when the patients present with various sub-types of lymphoma because of their peculiar heterogeneity in clinical presentation. The most probable chances of successful management or curing patients depend on early and accurate diagnosis. Most of the traditional tools they used like; biopsy and histopathological examination are invasive, long duration test procedures. This will reduce the time of diagnosis of lymphoma by using the combination of convolutional neural networks (CNN)-particularly modified ResNet and VGG models with the medical imaging into detecting and classifying lymphoma.The research is motivated to leverage the machine learning medical imaging technology advancement in developing a non-invasive, faster, and accurate diagnostic tool, which would be relied upon by clinicians to make better-informed decisions, eventually benefiting the patients with improved prognosis and reduced health costs. Transfer learning is the technique used to fine-tune the pre-trained weights of VGG16 and ResNet50 on the lymphoma dataset. The whole training phase is accelerated, and the performance of the model is also boosted when fine-tuned in this manner regarding the target task. Modification has been suggested in VGG16 and ResNet50 Models with pretrained weights to get optimum results on said dataset. The model is trained using optimization algorithms such as Adam and stochastic gradient descent (SGD), and its performance is evaluated using metrics like accuracy, precision, recall and kappa coefficient. The results indicate the effectiveness of Modified_VGG16-adam optimizer yielded 0.82 Precision, Recall 0.81,Kappa 0.7 and Accuracy 80.5 than VGG16 Models and Modified_ResNet50_adam optimizer gave 0.80 Precision, Recall 0.80,Kappa 0.58 and Accuracy 58.92 against ResNet50 Model. In future,other latest pretrained Models along with nature optimization algorithms can be integrated to improve overall accuracy