Liver tumour segmentation using contrast-enhanced multi-detector CT data: performance benchmarking of three semiautomated methods

Objective Automatic tumour segmentation and volumetry is useful in cancer staging and treatment outcome assessment. This paper presents a performance benchmarking study on liver tumour segmentation for three semiautomatic algorithms: 2D region growing with knowledge-based constraints (A1), 2D voxel classification with propagational learning (A2) and Bayesian rule-based 3D region growing (A3). Methods CT data from 30 patients were studied, and 47 liver tumours were isolated and manually segmented by experts to obtain the reference standard. Four datasets with ten tumours were used for algorithm training and the remaining 37 tumours for testing. Three evaluation metrics, relative absolute volume difference (RAVD), volumetric overlap error (VOE) and average symmetric surface distance (ASSD), were computed based on computerised and reference segmentations. Results A1, A2 and A3 obtained mean/median RAVD scores of 17.93/10.53%, 17.92/9.61% and 34.74/28.75%, mean/median VOEs of 30.47/26.79%, 25.70/22.64% and 39.95/38.54%, and mean/median ASSDs of 2.05/1.41 mm, 1.57/1.15 mm and 4.12/3.41 mm, respectively. For each metric, we obtained significantly lower values of A1 and A2 than A3 ( P < 0.01), suggesting that A1 and A2 outperformed A3. Conclusions Compared with the reference standard, the overall performance of A1 and A2 is promising. Further development and validation is necessary before reliable tumour segmentation and volumetry can be widely used clinically.