Automatic prostate cancer detection model based on ensemble VGGNet feature generation and NCA feature selection using magnetic resonance images

Prostate cancer is one of the most common types of cancer in men and its frequency is 28 per hundred thousand in the world. This cancer is detected using Magnetic Resonance Imaging (MRI). By using these images, an automatic prostate cancer diagnosis model must be introduced to simplify diagnosis process. A new MRI image dataset were collected from Firat University Hospital retrospectively. This image corpus contains malign and benign prostate cancer images. A novel transfer learning based model is presented for this dataset. Deep feature generation, feature selection with neighborhood component analysis (NCA) and classification are the primary phases of these model. (i) Deep features of the used prostate MRI images are extracted using VGG16 and VGG19 networks. These networks are pre-trained and they were trained on ImageNet dataset. Three fully connected layers (fc6, fc7 and fc8) of these networks are used to generate features and the generated features are merged. (ii) NCA selects top 500 features and (iii) the features chosen are classified using Cubic k nearest neighbors (kNN) algorithm. By deploying the presented ensemble VGG feature generator and NCA selector based technique, 98.01% accuracy was calculated. Moreover, other widely used performance evaluation metrics and confusion matrices were given to evaluate this model comprehensively. Results and findings obviously denoted the success of the recommended ensemble VGG feature generator and NCA selector based prostate cancer classification model.