Deep learning-based approach for the characterization and quantification of histopathology in mouse models of colitis

Soma Kobayashi, Jason Shieh, Ainara Ruiz de Sabando, Julie Kim, Yang Liu, Sui Y. Zee, Prateek Prasanna, Agnieszka B. Bialkowska, Joel Saltz, Vincent W. YangbioRxiv 2022 · DOI

TL;DR

This paper introduces a deep learning pipeline to automatically identify and quantify "involved" (diseased) and "uninvolved" (healthy) regions within histological images of mouse colons. This approach overcomes the limitations of manual pathological review, enabling more objective and efficient analysis of colitis models, which is crucial for understanding disease mechanisms and developing new treatments.

What Problem Does This Paper Address?

Inflammatory Bowel Disease (IBD) is a significant health issue, and researchers often use mouse models to study its pathogenesis and test new therapies. A key challenge in these studies is understanding the spatial distribution of disease within the colon. While advanced techniques like flow cytometry can analyze immune cells at a single-cell level, they often require processing whole organs, losing the crucial spatial context of where disease is occurring. Histological examination (looking at tissue under a microscope) can reveal this spatial information, but manually identifying and quantifying diseased versus healthy regions is time-consuming, subjective, and prone to inter-observer variability. This paper aims to bridge this gap by developing an automated, deep learning-based method to accurately segment and quantify these distinct histological regions in mouse colitis models.

Methods

The core of this work is a two-phase deep learning approach using Convolutional Neural Networks (CNNs), specifically the ResNet-34 architecture.

  • Phase 1: Colitis vs. Control Classification: Initially, a ResNet-34 model was trained to classify entire image patches as either "Colitis" or "Control" based on the overall status of the mouse. This provided a foundational understanding but didn't account for the heterogeneity within diseased colons.
  • Phase 2: Involved vs. Uninvolved Classification: To capture intracolonic heterogeneity, a more granular approach was taken. After extracting features from the Phase 1 model, k-means clustering was used to identify distinct patch classes. These classes were then used to generate more precise "Involved" and "Uninvolved" ground truth labels. A new ResNet-34 model was trained on these refined labels.
  • Small Patch Classifier: A separate, smaller ResNet-34 model was trained to classify 32x32 pixel patches into categories like 'Background', 'Muscle', 'Tissue', and 'Submucosa'. This was used to filter out irrelevant tissue areas (e.g., muscle, background) before the main analysis, improving the efficiency and focus of the pipeline.
  • Overlay Generation: The final model's predictions were visualized by overlaying "Involved" (red) and "Uninvolved" (green) regions onto the original histological images, providing a clear visual representation of the segmentation.
  • Mouse Model Prediction: Finally, the proportions of "Involved" and "Uninvolved" patches were used to train a Linear Discriminant Analysis (LDA) classifier, which could then predict the specific mouse model of colitis based on its histological profile.

Datasets

The study utilized an archived cohort of formalin-fixed paraffin-embedded (FFPE) swiss-rolled colons from three mouse models of colitis and their respective controls. These models included:

  • Klf5△IND: An inducible intestinal epithelium-specific knockout model.
  • DSS: Dextran Sodium Sulfate chemical induction model.
  • Combined Induction: A model combining both Klf5△IND and DSS treatments.

A prospective mouse cohort was also used to validate the robustness of the developed pipeline.

Key Insights and Ideas

  • Spatial Heterogeneity is Key: The paper strongly emphasizes that simply classifying a whole mouse colon as "diseased" or "healthy" misses critical information. The presence and distribution of both "involved" and "uninvolved" regions within the same colon are highly informative.
  • Deep Learning for Histopathology: This work demonstrates the power of deep learning, specifically CNNs, to automate complex and subjective tasks like histopathological analysis, offering objectivity and scalability.
  • Hierarchical Classification: The two-phase training approach (first broad "colitis/control," then fine-grained "involved/uninvolved") is a smart way to leverage existing data and progressively refine the model's understanding.
  • Patch-Level Analysis is Sufficient: By breaking down whole slide images into smaller patches, the model can effectively learn and classify histological features without needing to process the entire massive image at once.
  • Histological Signatures Differentiate Models: The study successfully shows that distinct patterns of "involved" and "uninvolved" histology can be used to differentiate between different mouse models of colitis, suggesting that these models recapitulate different aspects of human IBD.

Key Results

  • The developed "Involved" vs. "Uninvolved" classifier achieved an overall F1 score of 90.1% on the held-out test set.
  • The classifier trained with the refined, patch-level ground truth labels (Phase 2) showed improved agreement with pathologist-generated labels compared to the initial broad classification.
  • The LDA classifier, trained on the proportions of "Involved" and "Uninvolved" patches, could predict the mouse model with an impressive overall F1 score of 95.75% on the prospective cohort.
  • The study identified distinct histological "patch classes" (e.g., 'Crypts', 'Lightly Packed', 'Rosettes' for uninvolved; 'Inflammatory', 'Crypt Dropout', 'Crypt Dilation', 'Distorted Glands' for involved) that were differentially enriched across control and various colitis mouse models.

Limitations and Future Work

  • Mouse Model Specificity: The current application is limited to the specific mouse models included in the study.
  • Future Work:
    • Expand the training data to include a wider range of colitis models (e.g., TNBS, IL-10 knockout) to increase the method's generalizability.
  • Integrate immunohistochemical staining to map immune cell populations to the identified "involved" and "uninvolved" regions, providing a deeper understanding of cellular contributions to pathology.
  • Explore the potential of the "Small Patch Classifier" to further improve the performance of the "Involved" vs. "Uninvolved" model.
  • Investigate the clinical value and transferability of these histological phenotypes to human IBD.
  • Leverage the pipeline to analyze borderline regions and improve the grading of colonic histology.

Summary updated

Abstract

Abstract Inflammatory bowel disease (IBD) is a chronic immune-mediated disease of the gastrointestinal tract. While therapies exist, response can be limited within the patient population. As such, researchers have studied mouse models of colitis to further understand its pathogenesis and identify new treatment targets. Although bench methods like flow cytometry and RNA-sequencing can characterize immune responses with single-cell resolution, whole murine colon specimens are processed at once. Given the simultaneous presence of colonic regions that are involved or uninvolved with abnormal histology, processing whole colons may lead to a loss of spatial context. Detecting these regions in hematoxylin and eosin (H&E)-stained colonic tissues offers the downstream potential of quantifying immune populations in areas with and without disease involvement by immunohistochemistry on serially sectioned slides. This could provide a complementary, spatially-aware approach to further characterize populations identified by other methods. However, detection of such regions requires expert interpretation by pathologists and is a tedious process that may be difficult to perform consistently across experiments. To this end, we have trained a deep learning model to detect ‘Involved’ and ‘Uninvolved’ regions from H&E-stained colonic slides across controls and three mouse models of colitis – the dextran sodium sulfate (DSS) chemical induction model, the recently established intestinal epithelium-specific, inducible Klf5 Δ IND ( Villin-CreER T2 ;Klf5 fl/fl ) genetic model, and one that combines both induction methods. The trained classifier allows for extraction of ‘Involved’ colonic regions across mice to cluster and identify histological classes. Here, we show that quantification of ‘Involved’ and ‘Uninvolved’ image patch classes in swiss rolls of colonic specimens can be utilized to train a linear determinant analysis classifier to distinguish between mouse models. Such an approach has the potential for revealing histological links and improving synergy between various colitis mouse model studies to identify new therapeutic targets and pathophysiological mechanisms.
BibTeX
@misc{kobayashi2022deep,
  author = {Soma Kobayashi and Jason Shieh and Ainara Ruiz de Sabando and Julie Kim and Yang Liu and Sui Y. Zee and Prateek Prasanna and Agnieszka B. Bialkowska and Joel Saltz and Vincent W. Yang},
  title = {{Deep learning-based approach for the characterization and quantification of histopathology in mouse models of colitis}},
  year = {2022},
  doi = {10.1101/2022.05.12.491690},
  howpublished = {bioRxiv},
}
Figure