IDDF2022-ABS-0271 Deep learning-based classification distinguishes Crohn’s disease and intestinal tuberculosis from histopathological whole slide images

<h3>Background</h3> Crohn’s disease (CD) and intestinal tuberculosis (ITB) are both chronic granulomatous inflammatory disorders with broadly overlapping clinical, radiologic, endoscopic findings and pathologic, which might lead to misdiagnosis. This study was to identify useful deep learning models for distinguishing them from Whole slide images (WSIs). <h3>Methods</h3> The WSIs of surgically resected specimens were obtained from the First Affiliated Hospital of Sun Yat-sen University for establishing deep learning weakly-supervised models and internal testing. Additional WSIs were collected from the Sixth Affiliated Hospital of Sun Yat-sen University and The Second People Hospital of Foshan for external testing. The model performances were evaluated by area under curse (AUC) at slide-level and case-level respectively. Six pathologists reviewed the testing slides blindly, and the diagnosing results at the case level were used for comparisons with the model by Permutation Test. Heatmaps that highlighted areas of each patch for the classification were generated and Chi-square Test was used to calculate the statistical difference of histopathologic features. <h3>Results</h3> Overall, 2028 WSIs of 85 cases of CD and ITB were obtained in this weakly supervised model, with case-level AUC of 0.886, 0.892 and slide-level AUC of 0.954, 0.827 in the internal cross-validation and the external dataset, respectively. Compared with pathologists’ review, the diagnostic performance of the model (AUC=0.902) was better than junior pathologist1 (AUC=0.755, <i>p</i>=0.103) and junior pathologist2 (AUC=0.861, <i>p</i>=0.104), and was slightly inferior to senior pathologist1 (AUC=0.910, <i>p</i>=0.607) and senior pathologist2 (AUC=0.946, <i>p</i>=0.612) in the internal case testing. The model (AUC=0.876) was superior to the pathologist1 (AUC=0.782,<i> p</i>=0.182) and pathologist2 (AUC=0.821, <i>p</i>=0.391) in primary hospitals when testing the external dataset. Heatmaps were identified as some histopathologic features, such as granuloma or Langhans giant cells, muscular propria, and adipose tissue (<i>p</i>&lt;0.05). <h3>Conclusions</h3> We developed a deep learning diagnostic model for histopathological WSIs classification of Crohn’s disease and intestinal tuberculosis, which performed well and may help effectively improve the accuracy of clinicopathological diagnosis.