[Paper Review] Deep-Fill: Deep Learning Based Sinogram Domain Gap Filling in Positron Emission Tomography

Isaac Shiri, Peyman Sheikhzadeh+1 more|arXiv (Cornell University)|Jun 16, 2019

Medical Imaging Techniques and ApplicationsMedicine19 citations

TL;DR

This paper proposes Deep-Fill, a deep convolutional encoder-decoder network that reconstructs missing data in PET sinograms caused by inter-detector gaps. By training on augmented phantom data, the model predicts gap-free sinograms with high accuracy, significantly reducing streaking artifacts and improving image quality in reconstructed PET scans.

ABSTRACT

One of the major challenges in design and developing of PET, scanners are the presence of inactive areas between the detector blocks which degrade the image spatial resolution and leads to streaking artifacts especially when we employ analytical image reconstruction. The aim of this study is to assess the feasibility of generating the gap-free PET image using the deep convolutional encoder-decoder in sinogram space. The gap-corrupted sinograms of simulated HRRT scanner, sinograms without gaps as ideal/ground truth and predicted sinograms owing to the implemented our deep-fill method were quantitatively compared. In total, 1293 phantom images divided into three main sets of training 1000, 150 Validation, and 143 test set. The 1000 training images were augmented using affine transformations with various sub-transforms including rotation (rotate randomly), translation to 12000 Image with 6 frequencies of 2 main methods. The Deep-Fill architecture consists of an encoder and a decoder part, and it is composed of convolution operation, max pooling, ReLU activation, concatenation, and up convolution layers. The gap image is going through the network along with all possible paths then the gap-free image was generated by decoder part of the network. The quality of the generated images was quantitatively assessed by different quality metrics in both sinogram space and reconstruction images. We demonstrated that deep learning based approaches applied to inter-detector gap filling can recover the missing data in sinogram with high quantitative accuracy and have the potential to significantly improve the reconstructed image quality and prevent degradation of PET image quantification.

Motivation & Objective

• To address the degradation in PET image quality caused by inactive regions between detector blocks.
• To investigate whether deep learning can effectively reconstruct missing data in sinogram space.
• To develop a method that preserves image quantification accuracy while minimizing streaking artifacts.
• To evaluate the performance of a convolutional autoencoder in generating gap-free sinograms from corrupted inputs.

Proposed method

• A U-Net-like encoder-decoder architecture with convolutional, pooling, ReLU, concatenation, and transposed convolution layers was used.
• The network was trained on 12,000 augmented training images derived from 1,000 original phantom studies using affine transformations.
• Input sinograms contained artificial gaps mimicking inter-detector dead zones; the model predicted the corresponding gap-free sinograms.
• The training process used ground truth sinograms (without gaps) as targets for supervised learning.
• Data augmentation included random rotation, translation, and multiple frequency sub-transforms to improve generalization.
• Model performance was evaluated using quantitative metrics in both sinogram and reconstructed image spaces.

Experimental results

Research questions

• RQ1 Can deep learning effectively recover missing data in PET sinograms due to inter-detector gaps?
• RQ2 How accurately can a convolutional autoencoder reconstruct gap-free sinograms from corrupted inputs?
• RQ3 To what extent does gap-filling improve the quality of reconstructed PET images?
• RQ4 Does the proposed method preserve image quantification accuracy compared to conventional reconstruction?

Key findings

• The Deep-Fill model achieved high quantitative accuracy in reconstructing gap-free sinograms, significantly reducing streaking artifacts.
• Image quality metrics in reconstructed images showed substantial improvement after gap-filling using the deep learning approach.
• The model demonstrated robust generalization across diverse phantom configurations due to extensive data augmentation.
• The method preserved image quantification accuracy, indicating clinical viability for improving PET image reconstruction.