Development and technical validation of an artificial intelligence model for quantitative analysis of histopathologic features of eosinophilic esophagitis

• Main Authors: Luisa Ricaurte Archila (Author), Lindsey Smith (Author), Hanna-Kaisa Sihvo (Author), Thomas Westerling-Bui (Author), Ville Koponen (Author), Donnchadh M. O'Sullivan (Author), Maria Camila Cardenas Fernandez (Author), Erin E. Alexander (Author), Yaohong Wang (Author), Priyadharshini Sivasubramaniam (Author), Ameya Patil (Author), Puanani E. Hopson (Author), Imad Absah (Author), Karthik Ravi (Author), Taofic Mounajjed (Author), Rish Pai (Author), Catherine Hagen (Author), Christopher Hartley (Author), Rondell P. Graham (Author), Roger K. Moreira (Author)
• Format: Book
• Published: Elsevier, 2022-01-01T00:00:00Z.
• Subjects: article
• Online Access: Connect to this object online.
• Notes: 2153-3539; 10.1016/j.jpi.2022.100144
• Keywords: Artificial intelligence; Digital pathology; Eosinophilic esophagitis; Deep learning; EoE; Eosinophils; Computer applications to medicine. Medical informatics; R858-859.7; Pathology; RB1-214
• Source: Journal of Pathology Informatics, Vol 13, Iss , Pp 100144- (2022)

Background: In an attempt to provide quantitative, reproducible, and standardized analyses in cases of eosinophilic esophagitis (EoE), we have developed an artificial intelligence (AI) digital pathology model for the evaluation of histologic features in the EoE/esophageal eosinophilia spectrum. Here, we describe the development and technical validation of this novel AI tool. Methods: A total of 10 726 objects and 56.2 mm2 of semantic segmentation areas were annotated on whole-slide images, utilizing a cloud-based, deep learning artificial intelligence platform (Aiforia Technologies, Helsinki, Finland). Our training set consisted of 40 carefully selected digitized esophageal biopsy slides which contained the full spectrum of changes typically seen in the setting of esophageal eosinophilia, ranging from normal mucosa to severe abnormalities with regard to each specific features included in our model. A subset of cases was reserved as independent "test sets" in order to assess the validity of the AI model outside the training set. Five specialized experienced gastrointestinal pathologists scored each feature blindly and independently of each other and of AI model results. Results: The performance of the AI model for all cell type features was similar/non-inferior to that of our group of GI pathologists (F1-scores: 94.5-94.8 for AI vs human and 92.6-96.0 for human vs human). Segmentation area features were rated for accuracy using the following scale: 1. "perfect or nearly perfect" (95%-100%, no significant errors), 2. "very good" (80%-95%, only minor errors), 3. "good" (70%-80%, significant errors but still captures the feature well), 4. "insufficient" (less than 70%, significant errors compromising feature recognition). Rating scores for tissue (1.01), spongiosis (1.15), basal layer (1.05), surface layer (1.04), lamina propria (1.15), and collagen (1.11) were in the "very good" to "perfect or nearly perfect" range, while degranulation (2.23) was rated between "good" and "very good". Conclusion: Our newly developed AI-based tool showed an excellent performance (non-inferior to a group of experienced GI pathologists) for the recognition of various histologic features in the EoE/esophageal mucosal eosinophilia spectrum. This tool represents an important step in creating an accurate and reproducible method for semi-automated quantitative analysis to be used in the evaluation of esophageal biopsies in this clinical context.