Automated Localization and Segmentation of Left Ventricle in Cardiac MRI using Faster R-CNN

Convolution neural networks are good choice for localization and extracting ROI (Region of Interest) in MR images. But such algorithms are data hungry and need huge training dataset. Unfortunately, the publically available LV segmentation datasets are not large enough to cover all variations and dynamics of heart chambers. The accuracy of traditionally employed very deep neural networks tend to get reduce due to over-fitting on small datasets and, due to this factor, these algorithms are inefficient whenever the heart shape is outside learning dataset and there is a large bias in the segmentation. We tried to solve this problem by using a special kind of deep neural networks called faster R-CNN Inception-ResNet networks. Such networks are equipped with the power of inception as well as residual modules. The inception modules, in our architecture, help to learn information in image locally and globally whereas it also empowered with residual modules which are computationally less expensive because we can get higher accuracy in lower epochs. This fact makes our R-CNN architecture faster than traditional deep neural network architectures employed for LV localization and segmentation task. Also our R-CNN architecture can extract training features with multiple abstractions which is suitable for our problem as the publically available cardiac MRI image datasets are small, captured through different viewpoints and have inhomogeneous spatial and intensity level resolution along with inherent noise.