Deep learning assisted MRI guided attenuation correction in PET

Abstract

Positron emission tomography (PET) is a unique imaging modality that provides physiological and functional details of the tissue at the molecular level. However, the acquired PET images have some limitations such as the attenuation. PET attenuation correction is an essential step to obtain the full potential of PET quantification. With the wide use of hybrid PET/MR scanners, magnetic resonance (MR) images are used to address the problem of PET attenuation correction. The MR images segmentation is one simple and robust approach to create pseudo computed tomography (CT) images, which are used to generate attenuation coefficient maps to correct the PET attenuation. Recently, deep learning has been proposed and used as a promising technique to efficiently perform MR and various medical images segmentation. In this research work, deep learning guided segmentation approaches have been proposed to enhance the bone class segmentation of MR brain images in order to generate accurate pseudo-CT images. The first approach has introduced the combination of handcrafted features with deep learning features to enrich the set of features. Multiresolution analysis techniques, which generate multiscale and multidirectional coefficients of an image such as contourlet and shearlet transforms, are applied and combined with deep convolutional neural network (CNN) features. Different experiments have been conducted to investigate the number of selected coefficients and the insertion location of the handcrafted features. The second approach aims at reducing the segmentation algorithm’s complexity while maintaining the segmentation performance. An attention based convolutional encode-decoder network has been proposed to adaptively recalibrate the deep network features. This attention based network consists of two different squeeze and excitation blocks that excite the features spatially and channel wise. The two blocks are combined sequentially to decrease the number of network’s parameters and reduces the model complexity. The third approach has been focuses on the application of transfer learning from different MR sequences such as T1 weighted (T1-w) and T2 weighted (T2-w) images. A pretrained model with T1-w MR sequences is fine tuned to perform the segmentation of T2-w images. Multiple fine tuning approaches and experiments have been conducted to study the best fine tuning mechanism that is able to build an efficient segmentation model for both T1-w and T2-w segmentation. Clinical datasets of fifty patients with different conditions and diagnosis have been used to carry an objective evaluation to measure the segmentation performance of the results obtained by the three proposed methods. The first and second approaches have been validated with other studies in the literature that applied deep network based segmentation technique to perform MR based attenuation correction for PET images. The proposed methods have shown an enhancement in the bone segmentation with an increase of dice similarity coefficient (DSC) from 0.6179 to 0.6567 using an ensemble of CNNs with an improvement percentage of 6.3%. The proposed excitation-based CNN has decreased the model complexity by decreasing the number of trainable parameters by more than 46% where less computing resources are required to train the model. The proposed hybrid transfer learning method has shown its superiority to build a multi-sequences (T1-w and T2-w) segmentation approach compared to other applied transfer learning methods especially with the bone class where the DSC is increased from 0.3841 to 0.5393. Moreover, the hybrid transfer learning approach requires less computing time than transfer learning using open and conservative fine tuning.