Comparison Performance of Prostate Cell Images Classification using Pretrained Convolutional Neural Network Models

Abstract

Prostate cancer is the most common cancer in men in 2019. In that year, in the United States 174,650 men (20%) had prostate cancer and the remaining 696.32 men (80%) had other cancers (lung, bronchus) etc). In cancer diagnosis, there are several problems such as errors in reporting the diagnosis and the need for a long time. Artificial intelligence has long been known to facilitate the detection process, but a comparison analysis of the model is needed to get more optimal results. This study aims to compare the performance of two pretrained models (i.e. AlexNet and GoogLeNet). The data used is the image of prostate cells taken from a light microscope at the Universitas Indonesia (UI) Hospital. This study uses k-fold cross-validation to validate the accuracy of a model used. Performance evaluation of pretrained models is based on performance metrics: accuracy, precision, recall (sensitivity), specificity and f-score and running time in the testing process. The best accuracy is obtained by GoogLeNet with 99.63% and 97.74% and the lowest accuracy is obtained by AlexNet with 99.13% and 94.11%. During the training, AlexNet had a shorter time with 47 seconds than GoogLeNet with 112 seconds. In testing times, AlexNet was also faster with 0.307 seconds than GoogLeNet with 0.372. This research is expected to assist researchers (pathologists, physician assistants, etc.) in choosing the right architecture for the classification of prostate cancer images in terms of time and accuracy.