This section presents the results obtained and is divided into four parts. In the first part, we present the results obtained from training the proposed architecture using two optimizers and six different activation functions. In the second part, we present the results of the four different designs to enhance the performance of the network. In the third part, we present the results of the comparison of our architecture against state-of-the-art models. In the fourth part, we present the results of the comparison between our architecture and different architectures that were developed for histopathology image classification.
4.2.2. The Results of Different Designs
Four different designs were tested to check the performance of the network. The first design considers the network with two dropout layers in the classifier block, and the normalization layer before the activation layer. The second design is similar to the first design but without any dropout layers. The third design places the normalization layer after the activation layer and adds dropout layers in the classifier block. The fourth design is similar to the third design but without adding any dropout layers in the classifier block.
Table 4 and
Table 5 show the results for both the Adam and RMSProp optimizers.
The ReLU activation function has the lowest score compared to other activation functions across the four designs. The best performance of ReLU (with Adam optimizer) was obtained in the third design, followed by the fourth design. The worst score was by using the second design. The results with the RMSprop optimizer were poorer than the Adam optimizer. The best result was achieved with the second design, followed by the third design. The lowest result was obtained with the first design. Overall, the best result achieved by ReLU was 91.78%. The LeakyReLU activation function had slightly lower accuracy than the ReLU function. With Adam optimizer, the third design was the best performer, followed by the fourth design. The worse result was obtained with the second design. With the RMSprop optimizer, the best performance was obtained in the third design, followed by the second design. The worst performance was achieved by using the first design. The results of Adam and RMSprop optimizers were significantly different, especially for the first design. Overall, the best result achieved by LeakyReLU was 90.38%.
The ELU activation function had higher results compared to both ReLU and LeakyReLU. With Adam optimizer, the best performance was achieved by using the third design, followed by the fourth design. The lowest performance was achieved with the second design. With the RMSProp optimizer, the best result was achieved by using the fourth design, followed by the first design. The worst result was achieved with the second design. The results of both Adam and RMSprop optimizers were comparable, except for the second design. Overall, the best result achieved by the ELU activation function was 94.86%. The SELU activation function had slightly lower performance than the ELU activation function. With Adam optimizer, the best performance was achieved by using the second design, followed by the first design. The lowest performance was obtained with the third design. With the RMSprop optimizer, the best performance was achieved by using the second design, followed by the third design. The lowest performance was noticed with the first design. The results of Adam and RMSprop optimizers were significantly different, especially for the first design. Overall, the best result achieved by the SELU activation function was 94.25%.
The Tanh activation function achieved the highest results compared to all the other tested activation functions. With Adam optimizer, the best performance was obtained with the fourth design, followed by the third design. The lowest performance was achieved with the first design. With the RMSProp optimizer, the best performance was achieved by using the second design, followed by the fourth design. The lowest performance was obtained by using the first design. The results of both optimizers (Adam and RMSProp) were comparable. Overall, the best result achieved by the Tanh optimizer was 95.46%.
Using Adam optimizer, the first design was, overall, similar to the second design, meaning that the presence of the dropout layer did not increase the performance of the model. However, the performance of the third and fourth designs was higher, which indicates that the location of the normalization layer has an impact on the performance of the architecture. There are no significant differences between the third and the fourth designs, which indicates that the presence of the dropout layer does not increase the network performance.
Using the RMSProp optimizer, the performance of the first design was the lowest compared to the other designs. The second design achieved greater accuracy than the first design, which can indicate that the dropout layer can limit overfitting. The second, third, and fourth designs achieved a different performance, thus corroborating the hypothesis that the location of the normalization layer has a significant impact on the performance of the model. The third and the fourth designs were similar as well, indicating that the presence of the dropout layer has no effect on the model accuracy. All in all, based on the aforementioned results, we recommend practitioners to rely on the fourth design.
4.2.3. The Results over Benchmark CNN Architectures
In this section, the results of four benchmark CNN architectures are presented. Two sets of experiments were performed to compare our proposed architecture with four CNN popular benchmark architectures, namely, VGG16, VGG19, InceptionV3, and ResNet. The first set of experiments aims at comparing the architectures’ performance under the first design (dropout layer in the classifier block). All the original classifier blocks of the CNN were removed and replaced by two fully connected layers with a dropout layer after each fully connected layer, with a dropout probability of 0.5. The second set of experiments compared the performance of the architectures under the fourth design (without a dropout layer in the classifier block). Just like the first set of experiments, all the original fully connected layers were removed and replaced with two fully connected layers without any dropout layers. All the architectures were trained from scratch (i.e., no transfer learning was used).
In the first sets of experiments, using the first design, two optimizers were used as well. By using the RMSprop optimizer, the best performing architecture was our proposed architecture, followed by the VGG19 network. The worse performance was achieved by the InceptionV3 architecture. By using Adam optimizer, the highest performance was obtained by our proposed architecture, followed by VGG19. The poorest performance resulted from the ResNet architecture. Overall, our proposed architecture outperformed the other architectures taken into account. The results of the first set of experiments are shown in
Table 6. In the second set of experiments, using the fourth design, two optimizers were used as well. By using RMSprop, the highest performance was obtained by using our architecture, followed by VGG16. The lowest performance was achieved by using ResNet. By using Adam optimizer, the highest performance was achieved with our architecture, followed by VGG16. The lowest performance was obtained with the ResNet architecture. Overall, our proposed architecture achieved higher performance than the other tested CNNs. The results of the second set of experiments are shown in
Table 7.