Classifier Pipelines in HALO AI: Delivering Increased Efficiency for Sequential Classifier Workflows 

AI is helping to advance digital pathology research by streamlining workflows, increasing reproducibility, and enabling the collection of previously intractable data. Yet despite their capacity to improve analysis efficiency, deploying sequential AI classifiers in research can at times still be a highly manual process. With the release of HALO AI 4.0, we’re excited to introduce classifier pipelines, a new feature that streamlines the use of sequential classifiers. 

Classifier pipelines are a powerful functionality that allow users to daisy chain together AI models so that the output of one classifier is automatically fed into the next. Utilizing an intuitive graphical designer, developing complex classifier workflows in HALO AI is quick and easy. Once developed, pipelines can be run on their own or embedded within HALO modules, providing maximum flexibility. In this blog post we explore classifier pipelines—including their use in a familiar tool, their benefits, and how they are deployed—demonstrating how this capability can take classifier workflows to the next level in HALO AI. 

A Proven Example: SlideQC BF 

To illustrate a classifier pipeline in action, let’s consider SlideQC BF, our AI-powered tool designed for quality control of H&E- and IHC-stained whole slide images. SlideQC BF, which is available in HALO AI as well as being integrated in our HALO AP platform, detects the most common artifacts generated during slide preparation and scanning, such as air bubbles, dust, debris, folds, out-of-focus areas, and pen marks. 

Behind the scenes, SlideQC BF is constructed as chained classifiers. The process begins with a tissue detection classifier that feeds the detected tissue regions into a quality control classifier that detects and classifies the various artifacts. This pipeline automates the QC process and delivers a robust quality control workflow that significantly improves efficiency. 

The Benefits of Classifier Pipelines 

As SlideQC BF demonstrates, classifier pipelines offer significant workflow improvements, and with the capability to create custom pipelines in HALO AI users can now enjoy these benefits in diverse workflows. Fundamentally, classifier pipelines enable researchers to take a complex analysis task and break it down into specific, discrete, and simpler subtasks. This splitting of tasks enables greater optimization, with resolution and minimum object sizes adjustable for each classifier, and increased computing efficiency, resulting from some classifiers running at lower resolution and only on the selected output of earlier classifiers. 

Another efficiency benefit of classifier pipelines is the reduction in manual input required for sequential classifier workflows. Leveraging a pipeline, users no longer need to manually select the next classifier and region of interest after each step, freeing up time for other analysis or laboratory tasks. This also means that workflows can be more efficiently run during analysis downtime, such as overnight, without the need for regular monitoring and manual input to initiate the next classifier in the sequence. 

Further improving efficiency, the intuitive classifier pipeline designer simplifies the process of laying out an optimized sequence of classifiers and outputs / inputs. This feature assists users in getting the data they need as efficiently as possible, helping to avoid unnecessary classification steps or classifier reruns due to errors in workflow design. 

Lastly, classifier pipelines provide quantitative analysis results and interactive markups for every classifier in the pipeline, helping users explore the details of their analysis results. With data and images compiled under a single analysis job, users can more easily view classifier output in sequence, toggling class visibility on and off with interactive markups, and see the full analytic picture. These insights help ensure that each classifier in the chain is performing optimally and that the combined results provide the comprehensive analysis needed for complex projects. 

Leveraging Classifier Pipelines in HALO AI 

Whether a classifier pipeline will be embedded in a HALO module or run on its own, the process for creating a classifier pipeline begins in the HALO AI Classifier tab using the pipeline designer. Users add classifiers one at a time, selecting from custom or pre-trained classifiers, including Indica Labs’ new suite of HALO AI Apps. These pre-trained tissue classifiers and cell phenotypers enable researchers to accelerate their studies by reducing the amount of data and time required to train new models from scratch. 

In the pipeline designer, users also choose which class(es) from each classifier will be passed to the next step in the pipeline. For example, a researcher could string together SlideQC BF and the HALO AI Apps Ovarian H&E Tumor Tissue Detection and Pan-Cancer H&E Lymphocyte Cell Phenotyper to phenotype lymphocytes only in artifact-free ovarian tumor tissue.