Fusing Convolutional Neural Networks with Segmentation for Brain Tumor Classification

Early diagnosis is crucial for treating cerebral cancer. Significant progress has been made using noninvasive imaging methods combined with convolutional neural networks (CNN). Primarily, CNNs can extract relevant features from medical images, without requesting preliminary segmentations. However, additional ready-to-use information can be helpful for improving the accuracy of medical diagnostics. This paper compares different brain tumor classification methods based on CNNs. The proposed configurations consider an input-level fusion with two types of segmentation methods. One map is obtained via region-growing technique, from super-pixel partition; its role is to provide a ready-to-use simplified layout of the MRI scan. The other segmented image is a Class Activation Map (CAM), which incorporates the most relevant knowledge gained by a previously trained CNN. The experimental results indicate improved classification performance for the configurations that combine the original MRI scan with the segmented images into the input neural arrays.