Tittel: AI reconstruction in CT imaging
Sesjon: Medisinsk fysikk, torsdag 28. oktober kl 0830-1000
Moderator: Bente Konst, SIV
Rom: Agora 3, Clarion Hotel, Oslo
Background and techniques
Image reconstruction is an important part of the CT imaging chain. The traditional filtered back projection (FBP) results in acceptable image quality at moderate radiation doses. At low doses, image noise and artefacts increase. For the last decade iterative reconstruction (IR) has been the main method of reconstruction resulting in reduced image noise. But they also change the noise structure resulting in changes in image texture which gives these images a plastic looking appearance. With the advances in artificial intelligence, and especially deep learning, reconstruction algorithms can be developed by training on large data sets of clinical examinations. Today there are two deep learning algorithms on the market: AiCE (Canon Medical Systems) and True Fidelity (GE Healthcare). These have been trained on data sets with very different image quality characteristics; AiCE on IR images with improved spatial resolution and with different noise structure compared to FBP, and True Fidelity on high dose FBP images to keep the familiar texture of FBP.
Experience from our studies
Our initial studies on True Fidelity (TF) were promising regarding the potential for dose reduction while maintaining diagnostic image quality. In a clinical study we showed that images reconstructed with TF were perceived as superior compared to standard images with iterative reconstruction across a variety of quality criteria. Improved image quality could even be preserved when reducing slice thickness to 0.625 mm compared to the standard of 2.5 mm.
In a phantom study we explored the potential of TF for dose reduction regarding preserved low-contrast detectability. This is of high importance for the detection of e.g. small liver metastases. The dose reduction potential was estimated to 42% compared to iterative reconstructions and 55% compared to FBP.
On the background of these findings, we initiated a prospective clinical study with the primary objective to identify the potential for dose reduction for the detection of liver metastases (low-contrast detectability). The secondary objective was to determine the minimal acceptable dose level necessary to diagnose the cause of acute abdominal pain by evaluating image quality criteria. Results regarding the primary objective will be available shortly. For acute abdominal pain, even a dose reduction of up to 92.5% was still perceived as acceptable (dose reduction compared to our current standard protocol) and TF was superior to iterative reconstructions.