(Monday, 7/30/2018) 1:15 PM - 1:45 PM
Room: Exhibit Hall | Forum 5
Purpose: To improve the treatment of colorectal liver metastases (CLM) with ablation by utilizing fully convolutional neural networks and biomechanical, model-based image registration.
Methods: Thirty non-consecutive patients who had received treatment for CLM were retrospectively obtained for this work, 14 of whom had local recurrence at the treatment site. Manual contours of the liver, ablation probes, CLM margins, and ablation margin were created on both the pre-treatment contrast-enhanced CT images and the post-treatment contrast-enhanced CT images. A commercially available biomechanical, model-based deformable image registration (DIR) was then used to create a triangular element mesh on both image sets. This triangular mesh allowed the previously contoured CLM to be deformed onto the post-treatment imaging. A 3-dimensional assessment of minimum distance to agreement (DTA) between the deformed CLM to the ablation region was then performed to assess treatment efficacy. A fully-convolutional neural network was created in order to remove the necessity of manual contouring the liver. A blinded comparison between physician contours and manually generated contours of 26 patients showed: 52% (14/26) were preferred to the manual contours, and 92% (24/26) were deemed clinically useable as is or with minor (<1 minute) edits.
Results: There was a statistically significant difference in the minimum distance to agreement from the CLM to the ablation region between the 14 local recurrence patients and 16 non-local recurrence patients (p<0.01). Non-local recurrence patients received on average twice the ablation margin of those who did locally recur, 3.2 mm (range: 0-6.1 mm, median: 2.9 mm) and 1.1 mm (range: 0-5.6 mm, median: 0 mm), respectively.
Conclusion: The work presented here can assist in the treatment of CLM by rapidly assessing the treatment margins delivered. If the minimum distance to agreement is been shown to be insufficient, immediate steps can be taken to ensure necessary treatment.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
