Objective CT image-quality assessment remains difficult to scale when contrast-to-noise ratio measurement depends on manual region-of-interest placement. A 2026 investigation published in Insights into Imaging evaluated an open-source body and organ analysis framework for automated contrast-to-noise ratio analysis in chest CT. The work focused on contrast-enhanced computed tomography angiography and computed tomography pulmonary angiography, comparing automated measurements with manual assessments by three radiologists. The approach used automated segmentations of the aorta, pulmonary trunk and paraspinal muscles, then tested whether fat subtraction and binary erosion could reduce measurement deviations. The modified framework achieved agreement with expert measurements in internal and external validation cohorts, supporting reproducible quantitative image-quality assessment across chest CT examinations.
Automating Manual Image-Quality Measurement
Contrast-to-noise ratio is widely used because it is simple and clinically interpretable, but routine use remains constrained by manual region-of-interest placement. Manual placement takes time and introduces observer dependence, especially when measurements need to support large-scale evaluation, protocol optimisation or AI training pipelines. In chest CT, the need for reproducible metrics is particularly relevant because image quality affects lesion detection and accurate quantification.
The analysis included 100 contrast-enhanced chest CT scans acquired in 2022, with 50 computed tomography angiography scans and 50 computed tomography pulmonary angiography scans. The cohort had a mean age of 60.2 years, with 40% female patients. Scans came from five Siemens Healthineers CT systems and used a soft tissue reconstruction kernel with 1.0 mm slice thickness and increment.
Must Read: Optimising CT for Lung Screening
Three radiologists placed manual regions of interest in the ascending aorta, descending aorta, pulmonary trunk and autochthonous back muscles at the level of the pulmonary trunk. They placed the regions centrally and avoided vessel walls and surrounding or intersecting fat. The aortic contrast-to-noise ratio used the mean of the ascending and descending aorta measurements. Automated measurements used BOA segmentations of corresponding structures and calculated contrast-to-noise ratios from Hounsfield unit values and muscle standard deviation.
Segmentation Refinement Reduces Deviations
Unmodified BOA segmentations produced significantly lower contrast-to-noise ratios than manual measurements across both vascular structures and both protocols. Mean differences reached up to 6.30. The automated masks initially included vessel walls, adjacent fat in vascular masks, paraspinal fat in muscle masks and partial volume effects at muscle boundaries. These elements affected Hounsfield unit values and muscle standard deviation, leading to lower calculated contrast-to-noise ratios than manual regions of interest.
The refinement strategy applied binary erosion to aortic and pulmonary trunk masks and combined fat subtraction with binary erosion for autochthonous back muscle masks. Fat subtraction removed voxels between –200 and –40 Hounsfield units from the muscle mask. Binary erosion used kernel sizes from 2 to 10 voxels, with resampling to uniform voxel spacing before erosion so that erosion depth remained consistent across examinations.
These modifications progressively reduced deviations from manual measurements. For the aorta, mean BOA-examiner differences decreased from 3.84 with the original segmentation to 1.38 with m_erode8. For the pulmonary trunk, differences decreased from 6.28 to 2.36. Statistical significance compared with reader measurements disappeared once erosion depths of 4–6 were applied. Very deep erosions above 8 occasionally produced non-analysable segmentations in sarcopenic cases, where segmentation volume fell to zero.
External Validation Supports Reproducibility
A single preferred configuration combined m_erode6 with a_erode6 or p_erode6. This option balanced accuracy and robustness, avoiding insufficient correction at lower erosion levels and overcorrection at higher levels. Erosion level 6 became the first configuration in which both vascular targets showed no significant deviation from readers across computed tomography angiography and computed tomography pulmonary angiography.
External validation used The Cancer Imaging Archive. From 1,308 cases in the LIDC-IDRI dataset, automated contrast phase recognition identified 179 computed tomography angiography scans and 195 computed tomography pulmonary angiography scans. Random subsets of 50 cases from each protocol underwent independent radiologist review and comparison with the selected BOA configuration.
In the external validation cohort of 100 scans, the preferred configuration showed no significant difference from readers across all endpoints. Mean differences were 0.93 for aortic contrast-to-noise ratio in computed tomography angiography, 1.53 for aortic contrast-to-noise ratio in computed tomography pulmonary angiography, 0.77 for pulmonary trunk contrast-to-noise ratio in computed tomography angiography and 1.97 for pulmonary trunk contrast-to-noise ratio in computed tomography pulmonary angiography. Agreement reached excellent levels, with intraclass correlation coefficients of 0.89 for the aorta and 0.93 for the pulmonary trunk. Bland–Altman analysis showed minimal bias, with 0.16 for the aorta and 0.42 for the pulmonary trunk.
A minimally modified open-source segmentation framework can support fully automated, volumetric contrast-to-noise ratio assessment in chest computed tomography angiography and computed tomography pulmonary angiography. Fat subtraction in autochthonous back muscles and moderate binary erosion of vessel and muscle masks align automated measurements closely with radiologist assessments. The approach reduces observer dependence, supports scalable image-quality assessment and provides standardised quantitative metrics for protocol optimisation and AI-driven workflows. Further work remains needed for other regions, phases, reconstruction settings and image-quality dimensions beyond contrast-to-noise ratio.
Source: Insights into Imaging
Image Credit: iStock
References:
Beck N, Baldini G, Salhöfer L et al. (2026) Automated contrast-to-noise ratio analysis in chest CT: validation of an open-source segmentation approach. Insights Imaging, 17:88.
