Prostate Image Classification Using Pretrained Models: GoogLeNet and ResNet-50

Anatomic pathology errors often occur around 1% to 43% in diagnosis (depending on the disease). This error resulted in delayed treatment and an incorrect diagnosis. This study aims to detect the prostate cancer cells by using pretrained models of deep learning based on the value of performance metrics using a confusion matrix. Image data was taken from a light microscope at the Universitas Indonesia (UI) Hospital. In this study, 10-fold cross-validation was used as a validation of performance metrics on the model. In addition to the accuracy assessment, there is an assessment of precision, recall (sensitivity), specificity and F-score, as well as running time in the process. In this study, the accuracy value in each iteration has a relationship with other performance metric values, if the accuracy value decreases, the other performance metric values will also decrease, and vice versa. In training, ResNet-50 has an average accuracy value of 99.83%, 0.2% higher than GoogLeNet. Testing performances show that ResNet-50 has an average accuracy value of 98.02%, slightly higher (0.28%) than GoogLeNet. In F-score as well, GoogLeNet has a slightly smaller 0.54% difference than ResNet-50. It can be concluded that the pretrained models has good performances in prostate images classification.