Deep learning-based ablation margin assessment in thermal ablation of colorectal liver metastasis: preliminary results of an automatic segmentation workflow

<strong>Objective</strong> To work towards predicting treatment success during thermal ablation of colorectal liver metastasis (CRLM) by exploring the feasibility and accuracy of a deep learning (DL)-based intra-procedural segmentation workflow. <strong>Methodology</strong> A retrospective multicenter cohort of three academic institutions was used, consisting of 143 patients treated for 211 CRLM with 201 ablation zones. Reference manual tumor and ablation zone segmentations were previously obtained through a consensus read of two radiologists. The reference minimal ablation margin (MAM), an independent predictor of treatment success, was obtained from the manual segmentations after applying rigid co-registration to the intra-procedural pre- and post-ablation contrast-enhanced CT (CECT) scans. Two 3D nnU-Net segmentation networks were trained on 169 CRLM and 161 ablation zones, and tested on 42 CRLM and 40 ablation zones. During pre-processing, the CECT scans were multiplied with extended liver masks. The global (i.e., patient-based) and local (i.e., lesion-based) dice scores were reported. Test cases with a local dice score &ge; 0.70 were used for DL-based MAM quantification. The mean absolute difference in MAM and correspondence in categorization (&lt; 0 mm, 0 to 5 mm, &gt; 5 mm) between the DL-based MAM and manual MAM were reported. Additionally, the predictability of local tumor progression was tested with receiver operating characteristic (ROC) analysis. <strong>Results</strong> The global and local dice scores for the tumor model were 0.58 &plusmn; 0.30 and 0.60 &plusmn; 0.35, respectively. The model missed 8/42 (19%) tumors. The global and local dice scores of the ablation zone model were 0.86 &plusmn; 0.12 and 0.89 &plusmn; 0.12, respectively. MAM analysis was feasible for 22/42 (52%) tumors. The mean absolute difference between manual MAM and DL-based MAM was 0.5 &plusmn; 0.7 mm. The MAM categories corresponded for 20/22 (91%) tumors. ROC analysis, although not completely reliable due to the low number of cases, resulted in area under the curve (AUC) values of 0.63 and 0.62 for the manual MAM and DL-based MAM, respectively. <strong>Conclusions</strong> The automatic DL-based segmentation workflow was successfully used in a majority of patients for accurate MAM quantification. By improving the tumor segmentations’ quality, the feasibility of user-independent MAM analysis will increase further.