Interpretable AI Models for Pediatric Ulcerative Colitis Remission Prediction using Histopathology Images


Abstract:

Most children with UC present with extensive colitis at diagnosis, but the response to therapy is heterogenous. Identifying the optimal window for biologic treatment and risk stratification of patients remains an unmet need. The ‘Predictors of Response to Standardized Pediatric Colitis Therapy’ (PROTECT) study identified a combination of clinical and biologic predictors of clinical remission on mesalamine therapy alone at 52 weeks in treatment naïve patients. Our aim in this thesis is to leverage artificial intelligence approaches to histopathological classification of disease to enhance the current clinical predictive model. In this thesis, we utilized 100 hematoxylins and eosin (H & E) diagnostic treatment naïve mucosal biopsies from the PROTECT study. This thesis implemented a two step approach for slide preprocessing - stain normalization and informative patch selection (brightness), following with generating 512 x 512 patches. Project developed image analysis and machine learning workflow to discriminate UC patients who achieved corticosteroid (CS) free remission on 5-aminosalicylic acids (5-ASA) therapy alone for one year. Study first trained 23 machine learning (ML) classifiers on 221 handcrafted texture features (e.g., gray-level co- occurrence matrix (GLCM), entropy) for patch-level classification. Following, the top features were selected by the GINI index. Performance metrics included area under the receiver-operating-characteristic (AUROC) with 5- fold cross-validation. Study visualized the features with heatmap to highlight the region of interest to corporate with domain experts. As a result, 10399 patches from 100 patients (Male:55%; Age(y):13.5y(IQR:11-15); extensive disease:76%; CS free remission: 50%) were trained (80% training:20% test) on various ML classifiers. The extra tree classifier performed the best with a top mean accuracy of 80%±2% (compared with logistic regression, accuracy of 64% ±1%) and with an AUROC of 0.882 (95%CI:0.880, 0.887) on 5-fold cross-validation. Retraining on our top features (GLCM contrast, GLCM entropy with four angles and five displacement vectors (from 1px to 5px), histogram first order skewness) the extra tree classifier again performed favorably, with a mean accuracy of 70%±1% (logistic regression model accuracy of 63%±1%) and with a AUROC of 0.764 (95%CI: 0.757, 0.773). In conclusion, this study has demonstrated that standard-of-care pathology images and handcrafted features can be utilized to obtain an optimal ML-driven predictive model of pediatric UC disease course and aid in patient risk stratification. The next steps are to further develop model, incorporate predictive clinical covariates from PROTECT, and fusion the features with deep learning model’s feature.

End-to-end Interpretable Machine Learning Pipeline

Features Selection

Reference:


X. Liu, S. Prasath, I. Siddiqui, T. Walters, L. A Denson, PROTECT consortium, J. Dhaliwal. Machine learning-based prediction of pediatric ulcerative colitis treatment response using diagnostic histopathology. Gastroenterology, 166(5), 921-924, May 2024. doi:10.1053/j.gastro.2024.01.033 

Preprint at medRxiv doi:10.1101/2024.01.22.24301559


Posters/Abstracts/Presentations:

MS Thesis:

X. Liu. Interpretable Machine Learning for Histopathology Images Classification in Pediatric Ulcerative Colitis Remission Prediction. MS Thesis, Department of Computer Science, University of Cincinnati, USA, 2022.

Related Work:


X. Liu, J. Reigle, V. B. S. Prasath, J. Dhaliwal. Artificial intelligence image-based prediction models in IBD exhibit high risk of bias: A systematic review. Computers in Biology and Medicine, 171, 108093, March 2024. doi:10.1016/j.compbiomed.2024.108093


Back to Histopathology Projects Main Page. Back to Research Page.