Utilizing Deep Learning to Analyze Whole Slide Images of Colonic Biopsies for Associations Between Eosinophil Density and Clinicopathologic Features in Active Ulcerative Colitis

Abstract

Background Eosinophils have been implicated in the pathogenesis of ulcerative colitis and have been associated with disease course and therapeutic response. However, associations between eosinophil density, histologic activity, and clinical features have not been rigorously studied. Methods A deep learning algorithm was trained to identify eosinophils in colonic biopsies and validated against pathologists’ interpretations. The algorithm was applied to sigmoid colon biopsies from a cross-sectional cohort of 88 ulcerative colitis patients with histologically active disease as measured by the Geboes score and Robarts histopathology index (RHI). Associations between eosinophil density, histologic activity, and clinical features were determined. Results The eosinophil deep learning algorithm demonstrated almost perfect agreement with manual eosinophil counts determined by 4 pathologists (interclass correlation coefficients: 0.805–0.917). Eosinophil density varied widely across patients (median 113.5 cells per mm2, interquartile range 108.9). There was no association between eosinophil density and RHI (P = 0.5). Significant differences in eosinophil density were seen between patients with Montreal E3 vs E2 disease (146.2 cells per mm2 vs 88.2 cells per mm2, P = 0.005). Patients on corticosteroids had significantly lower eosinophil density (62.9 cells per mm2 vs 124.1 cells per mm2, P = 0.006). No association between eosinophil density and biologic use was observed (P = 0.5). Conclusions We developed a deep learning algorithm to quantify eosinophils in colonic biopsies. Eosinophil density did not correlate with histologic activity but did correlate with disease extent and corticosteroid use. Future studies applying this algorithm in larger cohorts with longitudinal follow-up are needed to further elucidate the role of eosinophils in ulcerative colitis.

Simplified Histologic Mucosal Healing Scheme (SHMHS) for inflammatory bowel disease: a nationwide multicenter study of performance and applicability

“…Furthermore, basal plasmacytosis is seen only in a minority of cases of relapse and thus is a relatively insensitive marker [ 15 ].⁠ Concerning point (c), there is only limited data on the association between eosinophils and relapse. Recently, Vande Casteele et al [ 16 ]⁠ quantified eosinophils in colonic biopsies and found correlations between eosinophil density and both disease extent and corticosteroid therapy, but crucially not with histologic activity. These factors confirm our belief that neutrophils are the only real marker of activity in IBD.…”

“…Because basal plasmacytosis is known to have high predictive value for the first diagnosis of IBD [ 7 , 8 , 11 , 12 ]⁠ and for the differential diagnosis with other forms of colitis [ 13 ],⁠ it has been hypothesized that an absence of basal plasmacytosis is required in healed IBD [ 14 ].⁠ This is contradictory, because the presence of basal plasma cells in this phase of disease is a sign of pre-existing IBD, and the number of plasma cells needed to document this feature is not known. Eosinophils, intermingled with plasma cells, are present at variable frequency in active and quiescent IBD [ 15 , 16 ].⁠ Therefore, the presence of eosinophils does not rule out inactive disease.…”

“…This is contradictory, because the presence of basal plasma cells in this phase of disease is a sign of preexisting IBD, and the number of plasma cells needed to document this feature is not known. Eosinophils, intermingled with plasma cells, are present at variable frequency in active and quiescent IBD [15,16]. Therefore, the presence of eosinophils does not rule out inactive disease.…”

Technological advances in inflammatory bowel disease endoscopy and histology

“…The first pioneering application of AI to histology in UC came from Vande Casteele et al (46) who trained a deep learning algorithm to quantify eosinophils in colonic biopsies. This AI model proved a high agreement with manual eosinophil count performed by human pathologists (ICC = 0.81-0.92) in a cohort of 88 UC patients with histologically active disease (46). Although no correlation between AI eosinophil count and histological activity was found, this work paved the way for other studies in the field.…”

Predictive models for endoscopic disease activity in patients with ulcerative colitis: Practical machine learning-based modeling and interpretation

“…It creates opportunities for exploring the relationships among features and building highly efficient models. Automated image recognition using deep learning methods also has been applied in the endoscopic images and pathological images recognition of IBD ( 21 , 26 – 28 ). Moreover, in search of new well-performing markers at the gene and microbiome level, the ML methods showed the greatest contribution in variables screening ( 29 , 30 ).…”