Diagnostic accuracy of pancreatic fibrosis grading with artificial intelligence: a single-center experience

Background: Chronic pancreatitis is a long-term fibro-inflammatory disease of the pancreas that worsens the quality of life of patients and is associated with life-threatening complications. Chronic pancreatitis is based on fibro-inflammatory changes in the tissue of the pancreas. The necessity of a more accurate identification of early fibrosis is related to discrepancies in assessments done with conventional visual scales. Aim: To evaluate the diagnostic accuracy of existing scales for visual assessment of the degree of pancreatic fibrosis compared to the techniques of whole-slide images (WSI) digital analysis. Methods: The morphometric analysis of the assessment of pancreatic fibrosis grade was carried out in the postoperative material from 118 patients with the use of two assessment semi-quantitative visual scales (those proposed by G. Klöppel B. Maillet in 1991 and by O. V. Paklina et al. in 2011), as well as with the use of the digital image marking in the ASAP software and an artificial intelligence algorithm in the QuPath software. Both intralobular and perilobular fibrosis grades were assessed, as well as their integrative index (II). Results: The assessment of pancreatic fibrosis with a 12-point rating system (II) according to Klöppel Maillett showed the following distribution: low grade fibrosis (LF), 55.9% of the cases, moderate fibrosis (MF), 25.4%, severe fibrosis (SF), 16.1%, no fibrosis (NF), 2.5% of the cases. According to the Paklina et al. scale, the results were as follows: LF, 66.1%, MF, 15.2%, SF, 12.7%, and NF, 5.9% of the cases. The comparison of the classical systems revealed a mismatch in the fibrosis grade in 27.9% of the cases; this resulted in the use of more objective methods for the WSI digital analysis. When comparing the average percentage of fibrosis area detected with ASAP with the fibrosis level according to Klöppel Maillett, the data were distributed as follows: with LF, the mean fibrosis area by ASAP was 6.9%, with MF, the fibrosis area was 30%, and with SF, 65%. Similarly, when counting cells in QuPath the following values were obtained: LF, 8.2%, MF, 27.9%, SF, 69.3%. The use of the WSI digital image analysis showed a higher concordance between them (R = 0.937) relative to the comparison between the classical rating scales (R = 0.811). When comparing the methods for assessment of fibrosis for the accurate diagnosis of chronic pancreatitis with ROC analysis, the best result was shown by the cell counting method with QuPath (AUC = 0.943). The method of area calculation with ASAP showed the AUC of 0.910, the classical Klöppel Maillett method demonstrated the AUC of 0.879 and the assessment according to Paklina et al. showed the AUC of 0.808. Conclusion: the WSI digital analysis has higher and significant concordance and diagnostic accuracy; therefore it can be used in clinical practice.