The Diagnostic Accuracy of Convolutional Neural Network Architectures for the Diagnosis of Brain Cancer

Background: There is a rising trend in the number of malignant tumors along with an increase in world population growth. More than 0.29 million cases of central nervous system tumors were diagnosed in 2018. The utilization of artificial intelligence for the histological analysis of tumors is being explored for the purpose of better patient care. Aim: The main purpose of the present research is to evaluate the diagnostic accuracy of two convolutional neural network architectures VGG 19 & ResNet18 for the histological examination of astrocytoma. Methods: After the ethical approval, the study is conducted on 190 anonymized digital pathology images which included 115 images of astrocytoma and 75 images of normal brain tissue. These images have been acquired from the anonymized H & E stained slides. From the total 190 images, one hundred twenty-eight (128) images are employed for the training set, 24 images for validation while 38 images for the test set. The digital images are classified by applying CNN architectures VGG-19 and ResNet-18. Results: The analysis of the test data revealed that the diagnostic accuracy and F1 score of ResNet-18 CNN architecture are 97% and 0.97 respectively while the diagnostic accuracy and F1 score of VGG-19 for the histological diagnosis of astrocytoma are 92% and 0.94 respectively. Conclusion: The diagnostic accuracy (97%) of ResNet-18 CNN architecture for the histological diagnosis of astrocytoma is better than the diagnostic accuracy (94%) of VGG-19.