top of page

DeepMirror for mouse brain image annotation

Automatic image annotation tools for biomedical have emerged as a game-changer, simplifying the laborious task of analysing and annotating extensive datasets. In this case study, we examine how we built an automatic mouse brain image annotation tool for uniQure that speeds up image annotation by 45x.

uniQure is a biopharmaceutical company that specializes in the development of transformative gene therapies for patients with genetic diseases, including both rare and more prevalent disorders. uniQure specialises in liver-directed and central nervous system disorders, making use of their best-in-class adeno-associated virus (AAV) delivery system.

What is AAV gene therapy?

AAV gene therapy utilizes a modified version of the adeno-associated virus as a delivery vehicle to introduce therapeutic genes into cells. AAV is a small, non-pathogenic virus that naturally infects humans but does not cause disease. The AAV viral particles are modified to remove the viral genes responsible for replication and pathogenicity, loaded with the therapeutic genes of interest and typically administered through injection. AAV viral particles enter the target cells, release the therapeutic genes, and integrate into the cell's genome (or exist as episomes, small, circular DNA molecules within the nucleus). These therapeutic genes are designed to compensate for genetic abnormalities, promote therapeutic effects, or modulate cellular functions to treat various diseases.

One of the advantages of AAV gene therapy is its relatively low immunogenicity and ability to establish long-term effects (gene expression) in target cells. However, there are still challenges to overcome, such as optimizing delivery efficiency, minimizing immune responses, and ensuring the long-term safety and efficacy of the therapy. Ongoing research and clinical trials are focused on addressing these issues to further improve the potential of AAV gene therapy as a treatment modality.

uniQure in clinical trials

uniQure is making great strides in advancing gene therapy to the clinic. Currently, the company has recently received FDA approval for HEMGENIX (etranacogene dezaparvovec), the first and only one-time gene therapy for hemophilia in the EU (HEMGENIX, uniQure). In addition, uniQure has one gene therapy candidate in Phase I/II clinical study, AMT-130, a one-time treatment gene therapy for Huntington’s disease (NCT04120493; AMT-130, uniQure); and at least 6 other disclosed pre-clinical candidates, one of which, AMT-260, received this month FDA clearance of Investigational New Drug Application (Programs pipeline, uniQure; AMT-260 INDA Clearance, GlobalNews).

Building an automatic mouse brain image annotation tool

One of the current challenges in the gene therapy field is target specificity. Gene therapies must only reach and exert an effect on the target tissue, without affecting any other tissues. To determine the specific location of gene expression and evaluate the impact of gene therapies on different organs, uniQure routinely carries out histological screenings of various organs, including mouse brain.

The brain has remarkably complex anatomy, where the precise identification of its diverse regions plays a crucial role. When it comes to pinpointing specific areas of the brain, labelling these images becomes a necessary step, often demanding labour-intensive and highly specialized manual work that can be quite time-consuming.

uniQure approached us to develop an AI-powered tool that could automatically label mouse brain tissues to accelerate the identification of the effects of a central nervous system (brain) gene editing therapy.

uniQure provided us with 200 unlabelled images belonging to histological coronal sections of mouse brain images. We manually annotated these images in a semi-automated manner and built a deep learning model that detects these regions in images. (Fig 1).

Brain image automatic annotations with DeepMirror take 45x less time than manual annotations
Figure 1: Building an automatic mouse image annotation tool. uniQure provided 200 unannotated images, which we semi-automatically annotated. We used these annotated images to learn from and built an automatic annotation tool that can annotate 8 regions of interest within 10 seconds per image.

Manual annotation of up to 8 brain regions in a single brain section requires between 5 and 10 minutes on average by a highly specialized, experienced annotator. Our automatic image annotation algorithm can automatically label all 8 regions of interest in a single brain section within 10 seconds (Fig 2a). A 45x reduction in annotation time for uniQure.

We calculated Jaccard index, a measure of how similar or different two images (Fig 2b), where 0 indicates no overlap or similarity between the images, and 1 indicates a perfect match or complete overlap between the images. Our algorithm can annotate images with a Jaccard Index of up to 0.84.

Graph showing how automatic image annotation with DeepMirror takes 45x less time than expert manual annotations.
Figure 2: DeepMirror accelerates image annotation by 45x. a. Manual annotation of up to 8 brain regions in a single mouse brain section takes on average 7.5 minutes (450s). Automatic labelling by DeepMirror takes 10s, a reduction in labelling time of 45x. b. Jaccard Index of AI-annotated images versus manual annotation. Jaccard Index measures the intersection of the union between AI and manual annotations, 0 indicating no overlap and 1 indicating perfect overlap.


The integration of AI in biomedical image analysis holds immense promise for improving diagnostic accuracy, reducing interpretation time, and enhancing patient outcomes, thereby shaping a future where AI becomes an indispensable tool in the field of medical imaging. In this case study, we have accelerated the analysis of biomedical images. Allowing researchers to rapidly annotate large volumes of images, enabling them to extract meaningful insights and drive scientific discoveries at an accelerated pace. Moreover, automatic, consistent, and standardized annotations eliminate potential human errors and inconsistencies, enhancing the overall quality and reliability of the research findings. With faster and more accurate image annotation, companies may gain a competitive edge, enabling them to explore novel avenues of study and make significant strides towards better therapies.

Do you have an application for DeepMirror? Contact us to get building at



Want to stay up to date with DeepMirror news?

Thanks for subscribing!

bottom of page