Artificial intelligence for analysing chest CT images

Study size, design and location

A retrospective study into 17 people with suspected COVID-19. Images were from European and South American centres and a public dataset.

Intervention and comparator

The segmentation accuracy of 12 algorithms (incorporated into icolung) was compared with radiologists. The segmentation accuracy of the AI algorithm was calculated as the dice coefficient (statistical method used to judge the similarity of 2 samples). COVID-19 status was confirmed by laboratory testing.

Key outcomes

Dice scores for lung segmentation, binary lesion segmentation and multiclass lesion segmentation were 0.982, 0.724 and 0.469, respectively. The AI algorithm performed binary lesion segmentation (identifying abnormalities) with an average volume error that was better than visual assessment by human readers. The algorithm performed least accurately in multiclass lesion segmentation (including identifying consolidation and ground glass opacity, which the authors note are important lesions in COVID-19).

Strengths and limitations

The study is very small (n=17) and published as a preprint (rather than as a formally peer reviewed paper). There is no information on the size or composition of the training datasets. A multicentre dataset was used and it is unclear how this generalises to an NHS setting. The authors note that a version of the software in this study is available as icolung in the US and Europe. It is unclear how the software in this study differs from icolung.

Study size, design and location

A retrospective study of 337 chest CT scans from 314 people (173 women, 164 men) with a total of 470 pulmonary nodules included (Murchison et al. 2019a and Murchison et al. 2019b). Images were from 1 UK regional healthcare database. Inclusion criteria were people aged between 50 and 74, who currently smoke or have a history of smoking, or are reported to have radiological evidence of pulmonary emphysema.

Intervention and comparator

The segmentation accuracy of Veye Chest was compared with 3 experienced chest radiologists. The segmentation accuracy of readers was calculated as the dice coefficient between each radiologist's segmentation and the segmentations of the others and subsequently averaged (inter-reader dice coefficient).

When looking at nodules visible on sequential scans, nodule registration from the AI was scored as either a true positive-pair if the detected registration was included in the nodule registration reference standard, or as a false positive-pair. The mean discrepancy between growth percentages determined by radiologists and AI alone was calculated.

Key outcomes

The software was able to successfully segment 95% of the total 428 nodules between 3 mm and 30 mm. The performance of the AI software for segmenting pulmonary nodules on chest CT was comparable with that of experienced thoracic radiologists.

The mean growth percentage of lung nodule pairs was similar between readers and by standalone AI.

Strengths and limitations

This study was done in a UK setting. The Fleischner Society's definition for pulmonary nodules was broadly used during this study. Training data were from people aged 50 to 74 in a registry of people who smoke. It is unclear how these data will apply to other patient populations.

The study is presented as a conference poster presentation. The study population is relatively small.