Share
Preview
 
Hi ,

H&E whole slide images are large and tissue appearance is diverse. Unlike methods to find mitoses or segment tissue types, pathologists cannot annotate which regions of the tissue are associated with patient outcome - at least not with any high degree of certainty.

I’ve written before about different cost functions for outcome prediction models. This time I want to discuss ways to model patient outcomes from the morphology of tumor and adjacent stroma.

Tumor and Stroma Morphology

Many survival models from whole slides start by segmenting tumor from non-tumor tissue and modeling only the tumor. However, properties of tumor-adjacent stroma may also play a role in outcomes.

Bhargava et al. studied both stroma and cancer cells [1]. They calculated a variety of features to capture stromal texture, connectivity of nuclei, and nuclear shape and orientation, among others.

They found that stromal features were able to predict prostate cancer outcomes on the African-American subset of their dataset and out-performed alternative prognostic methods.


Tumor and stroma morphology feature extraction [1]

Nuclei Diversity
Lu et al. took a different approach and looked specifically at morphological heterogeneity [2]. They quantified these variations within a tumor by capturing the degree of heterogeneity and diversity in the shape, size, and texture of cancer nuclei. They then demonstrated the ability of these features to predict survival from early stage non-small cell lung cancer.


Low vs. high risk of lung squamous cell carcinoma [2]

Tissue Patterns

But the morphology of cancer and of stroma are not the only aspects of tissue that can be modeled. Abbet et al. studied tissue patterns -- specifically, the interaction between different tissues [3].

Instead of segmenting tumor and stroma specifically, they focused on clustering tissue patches. They trained a self-supervised model to capture tissue appearance. Then they clustered patches using this representation. Finally, they calculated feature vectors from each image as cluster probabilities and cluster transition probabilities. With these features, a linear model was able to predict survival from colorectal cancer.


Sample clusters (left) and hazard ratios for clusters and cluster transitions (right) [3]

Recommendations

While models for prognostics often make use of weak supervision, they can also model specific aspects of tissue morphology or architecture. An increasing number of studies have associated stroma properties with patient outcomes. Explaining the factors contributing to a good or poor outcomes are critical for advancing our understanding of cancer.


Pathology ML Advisory Services - Filling Up Quickly

Is there a new project you’re starting or a challenge you’d like to get past?

I only have a few slots left for advisory services this fall. My ML Assessment & Roadmap can help you advance your project using the latest state-of-the-art techniques. Plus ongoing support will guide you through the implementation.

If advancing your machine learning algorithms for pathology images is a priority for you, just reply to this email and let me know you’re interested.

 
Fight cancer & climate change with AI
Distilling the latest research to help R&D teams implement better models and create an impact
Pixel Scientia Labs, LLC, 9650 Strickland Rd., Suite 103-242, Raleigh, NC 27615, United States

Email Marketing by ActiveCampaign