Purpose: Accurate definition of the resectability is of paramount importance for patients with pancreatic cancer (PC), yet the existing methods to determine resectability are descriptive rather than quantitative. Here, we aim to develop a novel object-space support vector machine (OsSVM) method based on contrast-CT to quantitatively characterize the degree of vascular involvement, which is the main factor determining the PC resectability.
Methods: On 107 contrast-CT PC scans (56 locally advanced (LA), 25 borderline resectable (BR) and 26 resectable (RE)), tumors and the surrounding vessels were manually delineated and fed into intra-subject OsSVMs for optimized tumor-vessel separation hyperplanes. Margins of the resultant support vectors were calculated as indicators of the overall misclassification level for each case. To characterize the degree of vascular infiltration, several metrics were developed based on spatial distributions of the margins (Pvi, Pmax, Pmean, and Pvar), and margin-histograms (Hmean, Hvar, Hskew, Hkurt, and Hener). Each metric was first evaluated as an independent classifier for groups. Elastic net (EN) regularized regressions were then applied in each classification test to select a set of significant metrics, which were further evaluated by ROC analyses.
Results: Among the nine metrics, Pvi and Pmax were the two best classifiers for LA and BR, yielding AUCs of 0.91 and 0.89, respectively. EN selected a combined set (Pmax, Pvi, Pmean, Hskew, and Hkurt) that improved the AUC to 0.95. For the classification of BR and RE, all metrics except Pvi yielded AUCs greater than 0.96. Four metrics (Pmax, Pvar, Pmean, and Hener) were selected by EN to create the combined set reaching an AUC of 0.98.
Conclusion: We introduced a novel quantitative OsSVM method for PC resectability evaluation on contrast-CT. The derived metrics successfully differentiated LA, BR, and RE with high classification capability. OsSVM thus may be used to refine the definition PC resectability.