Diagnostic Performance of Deep Learning Image Reconstruction in Low Dose CT for the Detection of Acute Abdominal Conditions

Background Computed Tomography (CT) has become an essential tool in modern clinical medicine. With widespread availability, a rapid increase in the use of CT imaging has been observed over the last decades. With the associated increase in radiation exposure, the potential increased risk for radiation-induced malignancy has become a public health concern. This is especially true for CT scans of the abdomen and pelvis which currently account for 50% of the collective CT dose. As the benefit of dose reduction in general is offset by deterioration of image quality, technological advances to reduce radiation dose without compromising image quality are aspired in clinical practice.

In CT-image reconstruction, filtered back projection (FBP) has been the dominant image reconstruction technique algorithm since the early 1970s, complemented by the first commercial iterative reconstruction (IR) algorithms in 2009.

A novel deep learning image reconstruction (DLIR) algorithm received clinical approval in 2019 (TrueFidelity, GE Healthcare, Milwaukee, WI). Other vendor-specific algorithms for deep learning image reconstruction are also emerging (AiCE, Canon Medical Systems, Otawara, Japan). As explained by a technical white paper, having been trained with high-dose and low-dose FBP datasets across phantom and patient cases, the DLIR algorithm strives to suppress image noise without compromising image quality. The use of deep learning image reconstruction has demonstrated potential for improved image quality and dose reduction without shifting noise texture.

For patients with acute abdominal conditions, CT of the abdomen and pelvis is considered the best first- or second-line diagnostic approach. For these patients a fast and accurate diagnosis is of great importance to avoid treatment delay and subsequent complications such as gastrointestinal perforation in case of appendicitis or diverticulitis. On the other hand, it is also important to avoid unnecessary surgical intervention and the related complications. A possible low-dose CT protocol must therefore provide a non-inferior diagnostic performance to facilitate fast diagnosis and avoid overtreatment and inconclusive examinations.

Promising results have been reported regarding low-dose CT examinations with model-based IR and dose reduction of up to 75-80%. However, with the introduction of DLIR even further dose reduction seems feasible. Our own results from an image quality perception study with DLIR indicate that a dose reduction of up to 92.5% compared to standard CT might preserve acceptable diagnostic image quality (yet unpublished work).