HALO AI

Trainable deep learning for segmentation, classification, and phenotyping in digital pathology without programming.

Overview

HALO AI is a trainable artificial intelligence toolset developed by Indica Labs, designed to advance discovery in digital pathology through deep learning-powered image analysis. Built for researchers and pathologists working with brightfield and fluorescence imaging, HALO AI addresses complex challenges in tissue segmentation, nuclear and membrane segmentation, cell classification, and phenotyping — all without requiring any computer programming or prior AI expertise.

Fully integrated with the HALO® and HALO Link platforms, HALO AI employs a straightforward train-by-example interface that allows users to define classes, train neural networks by drawing annotations, and apply trained classifiers across entire studies of whole slide images or selected regions of interest. The platform is suitable for a wide range of applications in oncology, translational pathology, immuno-oncology, and beyond.

Core AI Capabilities

AI-Powered Annotation: A point-and-click workflow enables rapid image annotation and training data development, optimized specifically for pathology data in both brightfield and fluorescence contexts.
Nuclear Segmentation: Multiple trainable AI-based nuclear segmentation networks are available, allowing users to optimize segmentation accuracy when nuclear morphologies vary across samples or staining protocols.
Membrane Segmentation: AI-powered membrane segmentation networks can be trained to accurately delineate cell membranes for specific assay requirements.
Tissue Segmentation: Trainable MiniNet and DenseNet networks classify tissue into user-defined classes, supporting high-resolution classifiers in both brightfield and fluorescence imaging.
Cell and Object Phenotyping: Users can quickly train a phenotyper to quantify cell types or objects of interest, including cells with complex morphologies such as neuronal cells, even in the absence of biomarker information.
Slide Quality Control (SlideQC): A trainable AI-powered quality control network detects common artifacts in H&E and IHC images to ensure data integrity across large studies.

Simple Three-Step Workflow

Define Classes: Establish the tissue classes, cell phenotypes, or segmentation targets relevant to your specific application.
Train Network: Draw annotations using the intuitive AI annotation tool to provide training examples; the neural network learns from these examples in real time with no programming required.
Apply Classifier: Deploy trained HALO AI networks within HALO modules for tissue classification, nuclear and membrane segmentation, and phenotyping, or apply them directly to whole slide image studies or selected regions of interest.

Managing Variability in Pathology Images

HALO AI is designed to handle extreme variability common in image analysis, including diverse morphologies, staining protocol-induced alterations, differences in tissue quality, and uneven staining.
Networks can be trained to deliver accurate segmentation and classification results across large studies with high morphological diversity.
HALO AI can be trained to work across vastly different stain types, including PAMS, Trichrome, H&E, and IHC, making it highly versatile for multi-stain research programs.

Advanced Features and Tools

Real-Time Tuning: Watch a network train in real time, toggle markups on and off to evaluate performance, add training data, or adjust parameters on-the-fly during the training process.
Probability Thresholding: Once a model is trained, a probability map can be used as an alternative output to a traditional mask; a heatmap visualization displays probability from lowest (blue) to highest (red) for each class.
Interactive Markups: Investigate results with interactive markup images that allow toggling individual populations on and off, combinable with probability thresholding for detailed validation exploration.
Deploy in HALO Modules: Trained HALO AI models can be deployed within HALO modules, enabling nuclear segmentation, membrane segmentation, tissue classification, cell or object phenotyping, and SlideQC within a unified analysis environment.
Model Interoperability: ONNX compatibility enables flexible import and export of HALO AI networks between machines, and supports launching and viewing results from integrated third-party AI apps directly within HALO or HALO Link.
Collaborative Development via HALO Link: Colleagues can be invited to a study to collaboratively contribute training data through the HALO Link image management platform.
Classifier Pipelines: Multiple classifiers or AI Apps can be connected into a single workflow pipeline, operable within HALO AI or within a HALO module.
Validation Metrics: A three-part Train–Validate–Test workflow provides quantitative metrics to rigorously assess classifier performance, with output metrics tailored to the type of network used.
GPU Optimization: Trained networks can be adapted for specific GPUs to accelerate analysis; training jobs can be queued by number of iterations or time, enabling overnight runs to maximize productivity and facilitate evaluation of multiple model variants.

Pre-Trained HALO AI Apps

Breast IHC Tumor Tissue Detection: Detects, segments, and quantifies tumor and other areas in hematoxylin and DAB-stained whole-slide images of breast cancer.
NSCLC IHC Tumor Tissue Detection: Detects, segments, and quantifies tumor and non-tumor areas in DAB-stained whole-slide images of non-small cell lung cancer.
NSCLC IHC Cancer Cell Phenotyper: Quantifies non-cancer cells, IHC-positive cancer cells, and IHC-negative cancer cells in NSCLC whole-slide images.
Breast IHC Cancer Cell Phenotyper: Detects, segments, and quantifies cancer cells and other cells in IHC-stained breast cancer images.
NSCLC H&E Cancer Cell Phenotyper: Detects, quantifies, and segments cancer cells from non-cancer cells in H&E-stained NSCLC whole-slide images.
CRC H&E Cancer Cell Phenotyper: Detects, segments, and quantifies cancer cells and other cells in H&E-stained colorectal cancer images.
Pan-Cancer H&E Lymphocyte Cell Phenotyper: Detects and quantifies lymphocytes across H&E-stained whole-slide images of multiple tumor types.
Gastric H&E Tumor Tissue Detection: Segments tumor, stroma, necrosis/other, and glass areas in H&E-stained gastric cancer whole-slide images.
HNSCC H&E Tumor Tissue Detection: Segments tumor, stroma, necrosis/other, and glass areas in H&E-stained head and neck squamous cell carcinoma images.
Ovarian H&E Tumor Tissue Detection: Segments tumor, stroma, necrosis/other, and glass areas in H&E-stained ovarian cancer whole-slide images.
Breast H&E Cancer Cell Phenotyper: Detects and quantifies cancer cells in H&E-stained breast cancer tissue whole-slide images.