Search

In this work, we investigated the ability of several Convolutional Neural Network (CNN) models for predicting the spread of cancer using medical images. We used a dataset released by the Kaggle, namely PatchCamelyon. The dataset consists of 220,025 pathology images digitized by a tissue scanner. A clinical expert labeled each image as cancerous or non-cancerous. We used 70% of the images as a training set and 30% of them as a validation set. We design three models based on three commonly used modules: VGG, Inception, and Residual Network (ResNet), to develop an ensemble model and implement a voting system to determine the final decision. Then, we compared the performance of this ensemble model to the performance of each single model. Additionally, we used a weighted majority voting system, where the final prediction is equal to the weighted average of the prediction produced by each network. Our results show that the classification of the two ensemble models reaches 96%. Thus these results prove that the ensemble model outperforms single network architectures. [ABSTRACT FROM AUTHOR]

Convolutional Neural Network (CNN) for medical image classification has produced satisfying work [11, 12, 15]. Several pretrained models such as VGG19 [17], InceptionV3 [18], and MobileNet [8] are architectures that can be relied on to design high accuracy classification models. This work investigates the performance of three pretrained models with two methods of training. The first method trains the model independently, meaning that each model is given an input and trained separately, then the best results are determined by majority voting. In the second method the three pretrained models are trained simultaneously as interconnected models. The interconnected model adopts an ensemble architecture as is shown in [7]. By training multiple CNNs, this work gives optimum results compared to a single CNN. The difference is that the three subnetworks are trained simultaneously in an interconnected network and showing one expected result. In the training process the interconnected model determines each subnetwork’s weight by itself. Furthermore, this model will apply the most suitable weight to the final decision. The interconnected model showed comparable performance after training on several datasets. The measurement includes comparing the Accuracy, Precision and Recall scores as is shown in confusion matrix [3, 14]. [ABSTRACT FROM AUTHOR]

Researchers often use theoretical models which provide a relatively simple, yet concise and effective way of modelling various phenomena. However, it is a well-known fact that the more complex the model, the more complex the mathematical description is. For this reason, theoretical models generally avoid large complexity and aim for the simplest possible definition, which although makes models mathematically more manageable, in practice it also often leads to sub-optimal performance. Furthermore, the data collected during the observations usually contain confounding factors, for which a simple theoretical model can not be prepared. Overall, mathematical models are usually too rigid and sophisticated, and therefore cannot really deal with sudden changes in the environment. The application of artificial intelligence, however, provides a good opportunity to develop complex models that can combine the basic capabilities of the theoretical models with the ability to learn more complex relationships. It has been shown [16] that with neural networks, we can build such models that can approximate mathematical functions. Trained artificial neural networks are thus able to behave like theoretical models developed for different fields, while still retaining their overall flexibility, which guarantees an overall better performance in a complex real-world environment. The aim of our study is to show our notion that we can create an architecture using neural networks, which is able to approximate a given theoretical model, and then further improve it with the help of real data to suit the real world and its various aspects better. In order to validate the functionality of the architecture developed by us, we have selected a simple theoretical model, namely the Kermack-McKendrick one [4] as the base of our research. This is an SIR [2] model, which is a relatively simple compartmental epidemic model, based on differential equations that can be used well for infections that spread very similarly to influenza or COVID. However, on one hand, the SIR model relies too heavily on its parameters, with slight changes in them leading to drastic overall changes of the S, I and R curves, and on the other hand, the simplicity of the SIR model distorts its accuracy in many cases. In our paper, by using the SIR model, we will show that the architecture described above can be a valid approach to modeling the spread of a given disease (such as influenza or COVID-19). To this end, we detail the accuracy of our models with different settings and configurations and show that it performs better than both a simple mathematical model and a plain neural network with randomly initialized weights. [ABSTRACT FROM AUTHOR]