KholsaArchit Khosla

Director, Machine Learning






NajdawiFedaa Najdawi, MD

Director of Pathology







Artificial Intelligence Applications in Inflammatory Bowel Disease: Quantifying the Mucosal Microenvironment of Ulcerative Colitis



Background: The inflammatory bowel disease (IBD) mucosal microenvironment is complex and likely contributes to both disease progression and resolution in IBD. Characterizing this environment by spatially quantifying the cellular and tissue composition, rather than limiting assessment to disease activity, can enable us to identify clinically relevant features that can better evaluate therapeutic efficacy and long-term outcomes. Although histologic remission (HR) is emerging as a treatment target, there are limitations to the existing histologic indices, including the lack of agreed upon definitions and the inherent variability that relates to human subjective assessment. PathAI has developed machine learning (ML) algorithms to characterize and quantify the cells and tissue regions within ulcerative colitis (UC) biopsies that have the potential to be more sensitive and reproducible compared with traditional methods.


Methods: Convolutional Neural Networks (CNN) were trained to segment tissue and detect cells in colon biopsies. 637 whole-slide images (WSI) of hematoxylin and eosin (H&E)-stained biopsies covering the spectrum of disease severity in UC were used. More than 130k region and point-based annotations were provided by the PathAI's network of board-certified pathologists, indicating relevant tissue regions (e.g., erosion/ulceration, crypt abscesses, epithelium with neutrophil infiltration, normal epithelium, granulation tissue, basal plasmacytosis) and cell types (e.g., neutrophils, plasma cells, lymphocytes, and eosinophils). Human interpretable features (HIFs) measuring area, count proportions and densities were extracted from CNN model predictions. To assess model performance, Spearman correlation was calculated between extracted features and consensus expert manual pathologist scoring of UC (Nancy Histology Index, NHI). A random-forest classifier model using 17 of the earlier generated HIFs was also trained against consensus pathologists NHI scores. 237 slides were used to directly predict NHI classification and evaluated on 80 held-out test slides. The random forest classifier model was also used to predict HR defined as NHI < 2.


Results: CNN predictions of cell types showed concordance with manual pathologist consensus labels. F1 scores for cell types (neutrophils, eosinophils, lymphocytes, plasma cells, and epithelial cells) model predictions were either equivalent to or within 1 standard deviation of the mean scores of individual pathologists. Model-generated quantitative HIFs measuring active inflammation and associated epithelial injury (cryptitis, crypt abscess, erosion, and ulceration) correlated with increasing disease severity and pathologist-assigned Nancy Histological Index (NHI) scores. Quantitative HIFs measuring features of chronic inflammatory cells discriminated unremarkable biopsies from those with chronic inactive colitis. These findings highlight the robustness of the model's performance. Using the random forest classifier, we were able to accurately predict NHI scores, with a weighted kappa (k= 0.93) and Spearman correlation (⍴=0.93, p<0.001) when compared to manual pathologist consensus NHI scores. We also had an accuracy of 0.94 to predict HR.


Conclusions: These results demonstrate the concordance of ML-based quantification of UC histology with manual pathologist scoring and highlight the potential of this approach to enable standardized, rapid, and robust quantitation of the mucosal microenvironment for improved evaluation of disease activity, drug response and long-term outcomes.



  1. Highlight how machine learning models can quantify cell types and tissue regions in UC biopsies directly from H&E-stained whole slide images: 'Explainable AI”
  2. Explain how machine learning model predictions can be used to extract “human interpretable features” (HIFs) that measure area and count proportions, densities, and spatial distributions of cells in the colonic mucosa
  3. Highlight how machine learning based evaluation can enable quantitative assessment of histopathology, providing rich data amenable to powerful statistical analyses. The potential use as an assistive device that could reduce intra- and inter-observer variability is currently being explored.




Archit Khosla is a founding engineer and currently a Director of Machine Learning at PathAI, a start-up that aims to transform the domain of pathology using machine learning and deep learning techniques. Archit has a master's degree in machine learning for the Georgia institute of technology.His 5 years of work at PathAI involves using Machine Learning to drive faster, more accurate detection and diagnosis of diseases such as Cancer, Inflammatory Bowel Disease (IBD), and Non Alcoholic Steatohepatitis (NASH). Archit is passionate about improving machine learning techniques to aid in the advancement of healthcare.


Dr. Najdawi is a board certified Anatomical Pathologist specialized in gastrointestinal, liver and endocrine pathology. She works at the intersection of Artificial Intelligence and Anatomical Pathology at PathAI in Boston, where she is Director of Pathology. She is using her pathology expertise to guide a multidisciplinary team with a mission to improve patient outcomes with AI-powered pathology.Dr. Najdawi completed her AP training and two fellowships in gastrointestinal and endocrine pathology at Brigham and Women’s Hospital, Harvard Medical School. She is certified to practice medicine in the USA, Australia, and Jordan. Having a special interest in education, she is on the educational committee of the DPA, and running an advanced digital pathology and AI elective, for pathology trainees at PathAI. She has multiple peer-reviewed publications and presented her work at national and international meetings. Her current research is focused on leveraging AI and digital pathology tools to empower precision pathology.

40+ Expert Presenters

View the PV22 agenda.


Register for #PathVisions22! 

October 16-18 | Las Vegas, NV


Support & Exhibit Opportunities

Trade Show Sold Out!