Impact of Adaptive Mean Filter as the Preprocessing Stage of Histopathological Image Classification of Breast Tumor Using Transfer Learning VGG16 for Various Magnifications

Abstract

This study assesses the impact of using an Adaptive Mean Filter (AMF) as a preprocessing stage for classification of breast tumor histopathological images at various magnifications. The histopathological image was converted from red-green-blue (RGB) into grayscale before AMF is applied. In this study, AMF was performed with kernel sizes of 3 × 3 and 5 × 5 pixels. The datasets were extracted using transfer learning VGG16 before being classified using Bagging classifier. To obtain unbiased performance of the model, stratified K fold cross-validation with K = 10 was used. The dataset was divided into K-equal-sized folds. For each fold, the model was trained on the remaining K-1 folds then evaluated on the held-out fold. This process was repeated K times, with each fold used once as the validation set. The accuracy of the model was then averaged over the K folds to estimate its generalization performance. The AMF with a kernel size of 3 × 3 pixels improves the multi-class classification accuracy for magnifications of 40× and 200×, resulting in accuracy increases of 0.20% and 0.89%, respectively. However, at a magnification of 100×, the model's performance decreases. While the use of AMF with a kernel size of 3 × 3 pixels did not raise the accuracy at magnification 400×, it resulted in a lower standard deviation by 0.24%. In binary-class classification, the use of the AMF with a kernel size of 3 × 3 pixels improves accuracy by 1.10% for magnification 40× and by 0.85% for magnification 200×. However, when implemented at magnifications of 100× and 400×, the AMF filter results in decreased performance. In conclusion, the use of the AMF with a kernel size of 3 × 3 pixels as a preprocessing stage for the histopathological image classification of breast tumor has shown to have a positive impact on the accuracy of multi-class and binary-class classifications for magnifications of 40× and 200×, but not for magnifications of 100× and 400×. The results also indicate that the use of AMF filter can reduce the standard deviation compared to without AMF for some magnifications. However, caution should be considered when applying the AMF filter, as it can decrease the model performance in some cases.