HALO AI Publications

Deep Learning-Based Annotation Transfer between Molecular Imaging Modalities: An Automated Workflow for Multimodal Data Integration

Alan M. Race, et al, Analytical Chemistry, 2021
With increasing demand to correlate data across multiple imaging modalities, Race and colleagues demonstrate a mechanism by which annotations can be generated on images from one imaging modality and transferred to a second image modality for data integration (and optionally, back again). Further, they perform this workflow on mass spectrometry images of a pancreatic cancer mouse model and hematoxylin and eosin-stained sections. HALO and HALO AI image analysis software was used to develop a DenseNet algorithm to classify and generate annotations for pancreatic ductal carcinoma tumor, non-neoplastic acinar tissue, and connective tissue on H&E-stained slides. This method for bidirectional transfer of image annotations may enable novel workflows in the future.

Maternal obesity during pregnancy leads to adipose tissue ER stress in mice via miR-126-mediated reduction in Lunapark

Juliana de Almeida-Faira, et al, Diabetologia, 2021
In this study, researchers set out to understand how miR-126-3, a microRNA found at increased levels in offspring of maternally obese mice, functioned in adipocyte metabolism. de Almeida-Faria and colleagues used proteomic approaches to identify a novel ER protein that is a direct target of miR-126-3 called Lunapark. HALO and HALO AI were used to train a DenseNet algorithm to selectively identify crown-like structures in H&E-stained fat tissue. Further, de Almeida-Faria and colleagues demonstrate that maternal obesity in mice leads to an increased risk of type 2 diabetes in offspring by targeting miR-126-3 regulation. Therefore, miR-126-3 is identified as a potential therapeutic target that could impact obesity and type 2 diabetes.

Tumoral PD-1hiCD8+ T cells are partially exhausted and predict favorable outcome in triple-negative breast cancer

Liang Guo, et al, Clinical Science, 2020
Prior to this publication it was known that dysfunctional PD-1hi CD8+ T cells infiltrated tumors, although it was unknown if this phenotype played a role in triple-negative breast cancer (TNBC). Guo et al set out to explore this phenotype in triple-negative breast cancer and using HALO and HALO AI demonstrated using both quantitative multiplexed immunohistochemistry and multispectral fluorescence imaging that PD-1hi CD8+ T cells were found in TNBC patient tissue biopsy core analysis but largely absent from peripheral blood. Molecular analysis of these cells revealed expression of biomarkers associated with T-cell exhaustion and the authors hypothesize that this cellular phenotype could be useful for future stratification and as a prognostic marker in TNBC patients as the presence of PD-1hi CD8+ T cells are associated with favorable outcomes. In addition, future research with this cellular phenotype may provide opportunity for therapeutic advancement in TNBC, a challenging subtype of breast cancer to treat.

Comparing Deep Learning and Immunohistochemistry in Determining the Site of Origin for Well-Differentiated Neuroendocrine Tumors

Jordan Redemann, et al, Journal of Pathology Informatics, 2020
Metastatic neuroendocrine tumors behave differently according to site of origin and it is important clinically to identify the primary site in order to identify an appropriate therapy. The site of origin in neuroendocrine tumors are challenging to identify based on H&E alone and can require an immunohistochemistry (IHC) panel. Redemann and colleagues evaluated the performance of HALO AI, a deep-learning convolutional neural network (CNN) on site of origin identification from a set of metastatic well-differentiated neuroendocrine tumors with known sites of origin and compared against IHC slides scored by pathologists. HALO AI was trained with H&E-stained tissue microarrays and was then evaluated against IHC analysis to identify pancreas/duodenum, ileum/jejunum/duodenum, colorectum/appendix, and lung. Results showed that HALO AI correctly identified the site of origin in 70% of cases and IHC correctly identified 76% of cases. As this was statistically insignificant, the authors conclude that a trained CNN can identify a site of origin from a well differentiated neuroendocrine tumor using morphology data alone with accuracy similar to that of IHC, the clinical gold standard.

Independent Prognostic Value of Intratumoral Heterogeneity and Immune Response Features by Automated Digital Immunohistochemistry Analysis in Early Hormone Receptor-Positive Breast Carcinoma

Dovile Zilenaite, et al, Frontiers in Oncology, 2020
This study by Zilenaite and colleagues evaluated the prognostic value of digital image analysis using HALO on analysis of hormone receptor positive breast cancer IHC biomarkers including ER, PR, HER2, and Ki67 combined with information on tumor heterogeneity and immune response. HALO AI was used for tissue classification to differentiate tumor, stroma, and background (necrosis, artifacts, glass). For quantitative analysis of breast cancer biomarker expression and localization, the Multiplex IHC module of HALO was used. The authors demonstrate that prognostic modeling in hormone receptor positive breast cancer is possible using the computational approach presented here. They also show that the addition of tumor heterogeneity data improved their prognostic model.

Advanced Prostate Cancer with ATM Loss: PARP and ATR Inhibitors

Antje Neeb, et al, European Urology, 2021
Researchers set out to evaluate the role of the ATM kinase in metastatic castration-resistant prostate cancer (mCRPC) with the long-term goal of improving molecular stratification in patients. HALO and HALO AI were used in the analysis of 800 ATM immunohistochemistry samples. Neeb et al detected ATM loss by IHC in 11% of their patient cohort which was associated with increased genomic instability but was not associated with a worse outcome. An in vitro model of ATM loss showed sensitivity to ATR and PARP inhibitors, which may be further investigated in future clinical trials of patients with ATM loss.

Two distinct immunopathological profiles in autopsy lungs of COVID-19

Ronny Nienhold, et al, Nature Communications, 2020
An international consortium of researchers characterized lung tissue from patients with COVID-19 using transcriptomic, histologic, and cellular analyses. Nienhold and colleagues report two phenotypes associated with lethal COVID-19 disease. One showed high levels of interferon stimulated genes in the lungs as well as limited lung damage and high levels of cytokines and viral loads. The second phenotype included severe lung damage with low levels of interferon stimulated genes, low viral loads, and high levels of CD8+ T cells and macrophages. As patients with the first phenotype die sooner, this highlights the need for biomarkers to classify COVID-19 patients and potentially guide treatment. HALO AI was trained using annotations from a pathologist to identify lung tissue and the resulting output was confirmed by pathology review. HALO was also used for quantification of immunohistochemistry analysis of CD3, CD4, CD8, CD20, CD68, CD123, CD163, and PD1.