In this study, we aim to develop and test an automated deep learning detection and segmentation software for non-small cell lung cancer (NSCLC) that can automatically detect and segment tumors on CT scans and thus reduce the human variation. We will assess the level of agreement between a group of radiologists, performing manual versus semi-automatic tumour segmentation. To do so, we will provide radiologists with two sets of CT scans. The first set will be segmented manually; the second one will be segmented using the automated software program.
Subsequently, we will use the inter- and intra-observer variance from the clinical study in a simulation or modeling study. We also compare the time needed and the consistency in segmentations by the software to medical doctors performance.
Reliability and Agreement study:
Primary tumours of 25 lung cancer patients will be delineated by 6 segmentation experts.
1. Assess agreement between automatic segmentation and radiologists' segmentation The primary tumours of 25 patients will be manually segmented by the radiologists and automatically by the the tool. The time needed to perform this task and the reproducibility of the segmentation will be recorded. The degree of overlap between the ROs and the automatic contour will be assessed pairwise using the Dice coefficient.
2. Delination of tumours by the experts, assisted by the software tool For another 25 patients, the experts will be provided with an automatic delineation, performed by the tool. They have the possibility to adjust and validate it. The time needed will be recorded. The difference between the mean overlap fraction in the first situation (manual delineation of experts) and the second situation (delineation of experts+ software tool) will be assessed, using a multi-observer Dice coefficient.
3. Assessment of intra-observer variance The experts will repeat the segmentation of the lung tumours after 2 weeks. They will repeat the manual segmentation (n=25) and the semi-automatic segmentation (n=25). This will make it possible to assess the intra-observer variance in both situations.
4. Qualitative assessment of the experts' preferences using an in-house developed visualization toolbox.
