부산대학교 도서관

Detecting the spread of cancer using digital images requires expertise from a pathologist, takes time and vulnerable to human error. In our work, we offer an approach to using a convolutional neural network (CNN) to predict the presence of cancer by using pathological images. We have developed a CNN framework, applied data augmentation then trained our model using a single GPU (Graphic Processor Unit) provided by Google Collaborator. We have used data published in the form of a Kaggle dataset. Datasets are small pathology images which are derived from Camelyon16, namely PatchCam. This database consists of approximately 200,000 images categorized into two classes: cancer and no cancer. We have split the data with a proportion of 70% for training and 30% for validation. We can report satisfactory results regarding the 200,000 images in the training and validation processes with accuracy reaches 0.92. The fl scores have found to be 0.94 for class no cancer and 0.91 for class cancer, while the AUC grade has 0.98. Besides the presentation of competitive results, we discuss how such automatic systems may improve the cognitive capabilities of clinical experts.