Session Item

CT image reconstructions are usually limited by the scan field-of-view (sFoV) (50 cm in our institution) which is not enough for patients with high BMI and/or using fixation devices.  An extended-field-of-view (eFoV) reconstruction using truncated data to estimate the patient geometry is already implemented in the CT software, but reconstructions using truncated data often result in imaging artefacts and have an unknown uncertainty. This study addressed the image quality by developing a novel deep learning-based reconstruction algorithm (HDeepFoV) and the uncertainty by developing a 3D printed phantom.  

HDeepFoV uses a convolutional neural network (CNN) to estimate the patient geometry and HU distribution even outside the sFoV. The training of the CNN was done based on patient images that were fully covered by the sFoV of the CT scanner. Those images were then virtually enlarged into the eFoV region and a virtual CT scan was simulated based on the enlarged images. Finally, the reconstructions of the virtual CT scan and the enlarged patient images served as input and ground truth for the training of the CNN.
The new HDeepFoV method was compared against current commercial state-of-the-art software HDFoV using a large 3D printed breast phantom based on patient anatomy with slots for the insertion of tissue-mimicking inserts so geometrical and HU accuracy were evaluated. Patient image reconstructions were qualitatively evaluated by medical physicists and physicians for different treatment sites.

HDeepFoV reconstruction for the breast phantom (Figure 1a) shows a superior geometrical accuracy (deviations < 5 mm) whilst HDFoV deviations were up to 25 mm (Figure 1b).  HDFoV accuracy varied significantly with the volume within eFoV and slice position whilst HDeepFoV showed a more consistent behaviour. HU values obtained using tissue-mimicking inserts showed similar results for soft tissue with HDeepFoV performing better for lung and bone inserts. All patient images reconstructed with HDeepFoV were considered superior in a qualitative evaluation regarding image quality and geometrical accuracy. HDFoV reconstructions showed high HU values (similar to bone) in regions of soft tissue near the edges of the sFoV which was not observed using HDeepFoV. In addition, a CT radiopaque placed on the skin of a patient was reconstructed inside the body in one slice whilst another marker was suspended in the air in another slice obtained with HDFoV.  The reconstructions obtained with HDeepFoV placed the markers in more realistic positions (Figure 2).

eFoV reconstruction is an important resource since there is no alternative to the use of truncation in the eFoV region.  However, it should be used carefully since there is ground truth for patients and results depend on several aspects such as the volume within eFoV. Our results obtained with phantom and patients indicate HDeepFoV is more accurate and resulted in better quality images than current commercial versions.