Developing AI-Based Tissue Classifiers for Digital Pathology 

Artificial intelligence (AI) is transforming digital pathology, and at Indica Labs’ Pharma Services we are at the forefront of using AI to help advance our customers’ biomarker studies. Our team has developed HALO AI classifiers for various tasks across numerous pathologies, leveraging our expertise in machine learning and histopathology. These skills are fundamental for training accurate, efficient AI classifiers, and in this blog post we discuss these tools and share guidance on their development. 

Tissue Classifiers: Mapping Tissues 

Tissue classifiers are designed to segment and delineate distinct tissue compartments within an image. Common classifiers include those that distinguish tumor from stroma or necrosis, as well as more specialized tasks like fibrosis and glomeruli detection. These classifiers allow researchers to analyze data separately for each compartment, delivering invaluable insights into dynamics across tissues. 

For simple tissue segmentation, we often rely on the Tissue Classifier Add-on, a random forest-based machine learning tool available in HALO®, which is well-suited for capturing whole tissue regions efficiently. We leverage HALO AI’s MiniNet for more granular, AI-driven classification that still allows for quick iterations while brainstorming workflows, thanks to its relatively minimal training input requirements. However, our go-to network for the most robust tissue classification is DenseNet V2—a deep learning neural network capable of accurately classifying multiple tissue types, or “classes,” within a given image. 

Training these tissue classifiers, especially those using DenseNet V2, demands hundreds or even thousands of example tissue regions. Careful attention must be paid to selecting representative regions across the entire study, and annotations must be both accurate and consistent to avoid confusing the deep learning algorithm. Additionally, the hierarchy of classes is critical for the model to correctly understand the relative importance of each annotation. Pharma Services has delivered many such tissue classifiers that consistently meet or exceed expectations, ensuring accurate, reliable data for our customers. 

Nuclear Segmentation Classifiers: Precision at the Cellular Level 

Accurately detecting nuclei is a fundamental step in cell-based HALO image analysis, allowing for specific analysis of this cellular compartment and serving as a reference or first step for other cell-based segmentation. While HALO’s traditional nuclear segmentation works well for standard nuclear morphologies, it can struggle with variations in nuclear brightness, shape, or in images with high background staining, leading to false-positive detections. 

Although the pretrained AI-based nuclear segmentation networks in HALO can help mitigate these issues, complex cases require custom-trained nuclear segmentation classifiers. These classifiers can be trained to better recognize unique nuclear morphologies and manage variability.  

Training an optimized AI classifier for nuclear segmentation again requires providing hundreds, if not thousands, of training regions to efficiently “re-teach” the underlying AI network. It is critical to provide examples with variable shapes, brightness, color, texture, and size. For areas with high background, this involves carefully separating out the true cells from background of the same intensity. Each slide and study are unique, and Pharma Services carefully optimizes nuclear segmentation for each project. 

Membrane Segmentation Classifiers: Detecting Cellular Boundaries 

Detection of membrane and cytoplasmic staining is another important step in cell-based image analysis, and several HALO modules allow for traditional detection of cytoplasmic and membrane stains. Using this approach, the cytoplasmic radius is entered manually or is determined by membrane detection, which relies on dye intensity or color deconvolution and optical density. Variability in staining can complicate this process, making a single, simple solution difficult to apply across studies. 

As with nuclear segmentation, HALO now includes pre-trained AI-based membrane segmentation networks for brightfield and fluorescence, which offer users increased segmentation accuracy across studies. Still, when staining variability is high or a study features a complex staining set, custom membrane segmentation classifiers using HALO AI prove essential. When training accurate and efficient membrane classifiers, Pharma Services follow the same general guidelines used for nuclear segmentation classifiers, ensuring a representative and extensive training set. By training a classifier designed to handle the specific staining characteristics in a project, we can offer the best membrane segmentation for complex samples. 

Get Started with Pharma Services 

Object Phenotypers: Categorizing Object Types 

Object phenotyping is a powerful capability found in HALO AI that allows for distinguishing between specific cell types or phenotypes or between objects more broadly. Phenotyping is a highly effective and flexible tool to quantify cells based on morphology when biomarker information is absent and can assist with quantification and characterization of cells with complex morphologies. Accurate nuclear or membrane segmentation is required for phenotyping, and any pre-trained or custom nuclear or membrane segmentation network can be embedded within the phenotyper.  

Common applications for object phenotypers include classifying tumor vs. stroma cells, immune vs. non-immune cells, single or double biomarker-positive cells, and cells with specific staining patterns, such as apical cells. Object phenotypers can even be combined with tissue classifiers within HALO’s analysis modules, allowing for the phenotyping of cells within specific tissue compartments.  

To make a robust phenotyper, you must provide hundreds to thousands of training examples to the network, with examples for each phenotype. Additionally, the training set should represent the variability inherent in each phenotype, whether that is biomarker positivity range or morphologic diversity. This level of precision is key to delivering the most accurate and informative results for your study. 

Measuring Classifier Quality 

One of the most critical steps in developing AI classifiers is measuring their performance, addressing the question “how good is this classifier?” Pharma Services routinely leverages quantitative quality metrics, such as precision, recall, and the related F1 score, to measure classifier performance. Importantly, different types of classifiers may call for different metrics, and users should ensure the metrics used and their application are appropriate.  Whether for tissue-level classification or cell/object-level analysis, properly measuring quantitative quality metrices helps ensure that the final AI model is performing at an optimal level. 

Conclusions 

Whether you’re segmenting tissues, detecting nuclei, or phenotyping cells, HALO AI has the tools to take your image analysis to the next level, and the Pharma Services team is ready to help ensure your project succeeds. With HALO AI and our expertise, we provide reliable, accurate data that meets the highest standards in the industry. 

If you’re interested in learning more about Pharma Services, check out our prior introductory and QC process blog posts, or reach out to us by clicking the button below to discuss how we can advance your biomarker studies. 

Get Started with Pharma Services