Noise immunity of deep learning and its application in histopatology

As a result of the fluctuations that occur during the image acquisition, the image pixel values change and the image turns into a noisy form. Noise makes it difficult to distinguish the features learned during training from each other and negatively affects the model performance. In this study, the immunity of deep learning models to label noise, attribute noise and both label and attribute noises was measured. In this context: by considering the Accuracy performance of the six popular deep learning models, it was tried to determine the model with better immunity at various noise levels. The MNIST database of handwritten digits and the Fashion-MNIST datasets were used for experimental studies. Image labels were randomly changed for label noise at 5%, 10%, 20%, 30%, 40%, and 50%, respectively. For attribute noise, random noise was added to the features as 5%, 10%, 20%, 30%, 40%, and 50%, respectively. Noise-added images were trained with popular deep learning models such as DenseNet-201, InceptionV3, Mobilnet-V2, Resnet50, VGG16 and VGG19 to measure their noise immunity. Accuracy was used as performance measurement metric. As a result of the experiments, it was observed that the immunity of deep learning models varied depending on the model architecture and the number of layers. The results obtained were the VGG16 model with number of medium layers and better immunity at various noise types and levels. In addition, ring cell cancer cells were detected in histopathological images with popular deep learning models. Early diagnosis is critical since the stage of the disease is determinant in the severity, treatment, and survival rate of cancer. Because the disease is discovered at a late stage, predictions about the patient's chances of recovery and the course of the disease are often poor. Diagnosing ring cell gastric cancer at an early stage will increase the chances of patients receiving appropriate treatment. VGG16, VGG19 and InceptionV3 deep learning models were used for cancer cell detection in histopathology images with region of interest (RoI). Fine-tuning strategy was applied by customizing the last five layers of the deep network models according to the target data. The parameters of accuracy, precision, recall, and F1-score were used to evaluate the model performance. Signet ring cell dataset taken from the competition "Digestive System Pathological Detection, and Segmentation Grand Challenge 2019" was employed. When compared to results of the DigestPath2019 Grand challenge ring cell gastric cancer competition, higher accuracy rates were obtained using deep learning models with accurate Region of Interest(RoI). VGG16 model exhibited a higher performance with accuracy of 95% and a F1-score of 95% among the models.