Liver Segmentation using Turbolift Learning for CT and Cone-beam C-arm Perfusion Imaging

Abstract

Model-based reconstruction employing the time separation technique (TST) was found to improve dynamic perfusion imaging of the liver using C-arm cone-beam computed tomography (CBCT). To apply TST using prior knowledge extracted from CT perfusion data, the liver should be accurately segmented from the CT scans. Reconstructions of primary and model-based CBCT data need to be segmented for proper visualisation and interpretation of perfusion maps. This research proposes Turbolift learning, which trains a modified version of the multi-scale Attention UNet on different liver segmentation tasks serially, following the order of the trainings CT, CBCT, CBCT TST -making the previous trainings act as pre-training stages for the subsequent ones -addressing the problem of limited number of datasets for training. For the final task of liver segmentation from CBCT TST, the proposed method achieved an overall Dice scores of 0.874±0.031 and 0.905±0.007 in 6-fold and 4-fold cross-validation experiments, respectively -securing statistically significant improvements over the model, which was trained only for that task. Experiments revealed that Turbolift not only improves the overall performance of the model but also makes it robust against artefacts originating from the embolisation materials and truncation artefacts. Additionally, in-depth analyses confirmed the order of the segmentation tasks. This paper shows the potential of segmenting the liver from CT, CBCT, and CBCT TST, learning from the available limited training data, which can possibly be used in the future for the visualisation and evaluation of the perfusion maps for the treatment evaluation of liver diseases.

Liver Segmentation in Time-resolved C-arm CT Volumes Reconstructed from Dynamic Perfusion Scans using Time Separation Technique

“…Turbolift learning trains a multi-scale attention UNet in different stages, where the earlier training stages act as pretraining for the subsequent stage -assisting the model to learn from a small training dataset. The original Turbolift [7] stages were CT, CBCT, and CBCT TST. Here, the CBCT TST implied the reconstructed first coefficient (FCR) by the means of TST.…”

“…Liver segmentation in CT reconstructed volumes Previous research [7] only utilised the first coefficient for training since the associated basis function is a constant one and therefore the most similar to the classical reconstructed CT volumes (i.e. straightforward reconstructions).…”

“…Deep learning typically requires large training datasets that are comparable to the final test set. Turbolift learning [7] was proposed to combat this problem by employing a finetuning-based approach -using transfer learning in different stages. Turbolift starts by pretraining a modified version [10], [11] of the Multi-scale Attention UNet [12] on a dataset of healthy humans -the CHAOS dataset [10].…”

“…Similarly, the output of the model is compared at different scales as well -known as deep supervision [15] or multi-scale supervision [16]. The final output of the model (output from the final decoding block), along with three more outputs from the inner decoding blocks, are compared against the original ground-truth (used in [7]), as well as three downscaled versions of the ground-truth, respectively.…”

“…However, none of these offered completely automated liver segmentation in CBCT volumes. Turbolift learning has been proposed [7] to segment the liver from differently reconstructed CT and CBCT perfusion volumes, including TST.…”