AI4Life

AI4Life_Logo
Menu

Category: News

Categories
News

Second AI4Life Open Call: Announcement of selected projects

Second AI4Life Open Call: 
Announcement of selected projects

by Joran Deschamps, Florian Jug & Beatriz Serrano-Solano

The second AI4Life open call received 51 applications from diverse scientific disciplines (see this online report for more details).

In contrast to our first open call, the selection procedure has slightly evolved. We are now distinguishing two phases, a consultancy phase, and a project execution phase. The motivation for this evolution stemmed from the observation that we can maximize the value of our service to the scientific community by providing feedback (e.g. feasibility, existing tools, approaches we would take) to a larger number of projects, and then taking on a small number of projects for more in-depth support (similar to last year’s first open call) based on informed opinions of experts regarding the readiness of the project.

The selection of projects, for both phases, was conducted by an international group of project reviewers, similar to the procedure of last year.

Projects selected for Phase 1: Project Consultation

  1. Automated image analysis for mouse embryos phenotyping (#1)
  2. Analysis of necrotic cores in diabetes (#3)
  3. Bring isotropy to Volume EM (#7)
  4. Automated species annotation in turbid waters (#11)
  5. Segmenting dynamic 3D cell shape in crowded epithelia (#12)
  6. Optimizing calcium image acquisition with machine learning denoising algorithms (#13)
  7. Diatom symbioses at planetary scale (#14)
  8. Automated segmentation of actin filaments in intact cells (#21)
  9. Automated and Integrated Tunneling Nanotube (TNT) Detection and Counting (#22)
  10. Enhancing the effective resolution of intravital microscopy with digital expansion microscopy (#23)
  11. Automated myelin segmentation (#24)
  12. Ultrastructural protein mapping through Correlation of Light and Electron Microscopy (#26)
  13. The speed of life in trees – linking wood anatomy with lifespan and respiration (#28)
  14. Systematic quantitative characterization of mechanisms controlling spindle and nuclear scaling in tissue (#29)
  15. CKit- Mast Cell Tumor (#30)
  16. Novel regulators of chromatin response to DNA damage (#32)
  17. Scaling of actin networks with cell size (#34)
  18. NeuroScan – A 3D human motor neuron disease platform for high throughput drug screening  (#38)
  19. Cell Fate of Pax7-tdtomato transplanted cells in immune dysregulation during aging (#46)
  20. Correlating bacteria images from STORM and AFM microscopies in a combined analysis using registration methods (#51)

As the titles suggest, consulted projects covered a wide range of disciplines and topics. Consultation sessions were typically conducted with between two to four AI4Life experts and consisted of a one-hour discussion during which the experts could ask additional questions regarding the project and then provide opinions regarding the feasibility of the projects. Along with this kind of “reality check”, the experts also aimed at giving tangible advice on existing tools and methods that might help the consulted researchers in improving their workflows and achieving the desired analysis results.

Projects selected for Phase 2: Deep Learning project support

With the additional information from the consultancy phase at hand, the group of AI4Life experts came together and discussed the most promising projects to be taken on in phase 2 (deep-learning support). This decision was guided by a combination of project feasibility, readiness, re-usability by the community, and how well each project fits within the experiences and capabilities of the AI4Life team itself. These considerations were combined with the intention to maximize the utility and impact on the bioimage analysis community of this Open Call.

  1. Automated species annotation in turbid waters (#11*)
  2. Optimizing calcium image acquisition with machine learning denoising algorithms (#13)
  3. Diatom symbioses at planetary scale (#14*)
  4. Automated segmentation of actin filaments in intact cells (#21)
  5. Enhancing the effective resolution of intravital microscopy with digital expansion microscopy (#23)
  6. Ultrastructural protein mapping through Correlation of Light and Electron Microscopy (#26)
  7. NeuroScan – A 3D human motor neuron disease platform for high throughput drug screening (#38)
(*) Project goals were readjusted after consultancy between applicants and the AI4Life team.

Details about the reviewing procedure

Reviewing procedure

After assembling a panel of 17 international reviewers, we distributed anonymized projects among them. All personal and institutional information was removed, only leaving project-relevant data to be reviewed (including a short presentation and data examples). We aimed at receiving 3 independent reviews per project, requiring each reviewer to judge 9 projects in total.

Here is the list of questions we asked our reviewers via an electronic form:

  1. Please rank the following statements from 1 (Likely not) to 5 (Likely):
    1. The project has well-defined goals (and these goals are the correct ones).
    2. The proposed project will benefit from Deep Learning methods/approaches/tools.
    3. A complete solution will be purely deep learning (no intensive classical routine necessary).
    4. The project requires developing new approaches or tools.
    5. Approaches/scripts/models developed to solve this project will likely be reusable for other, similar projects.
    6. The project, once completed, will be useful for a broader scientific user base.
    7. The project’s current analysis pipeline can easily be improved.
    8. The project cannot be solved by simply offering one or two consultations to the users.
    9. The project seems feasible in a few months (of part time work).
    10. The project, once completed, will be interesting to computational researchers (e.g. within a public challenge).
    11. The applicants will have no problem sharing their data publicly.
  2. What are the key sub-tasks the project needs us to improve?
  3. How long would you expect it to take (in person/days) to generate sufficient training data (if applicable)?
  4. Once there is enough training data available, what workload do you anticipate in developing a viable solution for the proposed project and why?
  5. What is your estimated minimum number of days for successfully working on this project?
  6. What is your estimated maximum number of days for successfully working on this project?
  7. Do suitable tools for solving this task exist? What would you use?
  8. On a scale from 1 to 10, how enthusiastic are you about this project?
  9. Would this project qualify as an AI4Life project?
  10. Would this project qualify for Consultation?
 Scoring projects according to reviewer verdicts

We first aggregated all reviews per project by averaging numerical values and concatenating textual evaluations. We then computed two different aggregated scores: consultation and support based on the reviewer evaluations:

Consultation score
  • [0.8] Solvable in 1-2 consultation sessions (= 5 – “Not solvable”)1
  • [0.2] Solution likely reusable
Support score
  • [0.2] Interesting for the computational community
    • [0.1] Enthusiasm2
    • [0.1] Interesting for challenge
  • [0.8] Quality & Feasibility
    • [0.2] Quality of the application
      • [0.1] Well-defined project
      • [0.1] Pipeline can be easily improved
      • Persons/day to generate training data (tiebreaker
    • [0.2] DL suitable
      • [0.15] DL applies
      • [0.05] only Deep Learning and not classical image analysis
    • [0.2] Impact
      • [0.1] Impactful in the broader community
      • [0.1] Solution likely reusable
      • Workload to find a solution with training data (tiebreaker
    • [0.2] Time effort
      • [0.1] Feasible in a few months
      • [0.05] Min days3
      • [0.05] Max days
1 “Solution likely reusable” calculated from “Not solvable in 1-2 consultations” (scale inversion)
2 Enthusiasm rescaled from 0-10 to 0-5
3 Min and Max days rescaled to linearly decrease from an optimum of 60 days (score = 5)
 Consultation phase
  1. After ranking the projects, we decided on a scoring threshold for both consultation and project scores so that we could provide about 20 hours of consultations. We invited all applicants whose project scored either >3 on the consultation score (11 projects) or >3.8 on the project score (12 projects), to a 1h consultation with a panel of typically two to four suitable experts. Two projects declined the invitation, and one project scored high on both scores, resulting in 20 consultations.
  2. After each consultation, the experts drafted a summary of their suggestions to the scientists, and their opinion regarding the feasibility and readiness of the project, in particular with respect to the AI4Life Open Call time frame.
  3. A second consultation was offered to a subset of the projects (7). While this was not initially planned, we decided that those projects would benefit from additional feedback and guidance within the time frame of a single additional consultancy session.

Final decisions were made during a general assembly of reviewers and AI4Life Experts who would later take on the selected projects. The decisions suggested during this meeting were brought to the Open Call Selection Committee, which checked the soundness of all decisions and unanimously accepted the phase 2 project selections.

Who was involved in the process?

And now? What’s next?

The applicants whose projects were selected are now being contacted by the AI4Life team. All other projects that expressed interest in being part of the AI4Life Bartering Corner will be contacted to prepare publication on the AI4Life website, so computational experts can reach out to the scientists behind any of these projects and start a collaboration. 

If you did not have the chance to apply to the second Open Call, there will be a third and last call at the end of 2024. Subscribe to our newsletter, and you will be informed about the opening of the next call!

If you need help quicker, we recommend Euro-BioImaging’s web portal, where you can access a network of experts in the field of image analysis. Please note that this service may involve associated costs, but access funds for certain research topics are available through some initiatives. We and Euro-Bioimaging will be happy to help you find suitable opportunities!

Categories
News

Insights from the AI4Life team: Joran Deschamps and Craig Russell on AI4Life

Insights from the AI4Life team: Joran Deschamps and Craig Russell on AI4Life

by Beatriz Serrano-Solano

We are excited to share two interviews featuring Joran Deschamps, image analysis researcher and research software engineer coordinator at Human Technopole; and Craig Russell, AI engineer at EMBL-EBI, who discuss their views on AI4Life and its impact on the field of bioimaging and AI research.

Discover more about AI4Life through the eyes of these experts.

Categories
News

Interview with Joanna Hård, CEO of Amun AI

Interview with Joanna Hård, CEO of Amun AI

by Beatriz Serrano-Solano


We’re excited to announce an interview with Joanna Hård, CEO of Amun AI, where she discusses the scope of this innovative spin-off from AI4Life and KTH.

Amun AI’s flagship product is a plug-and-play version of the Hypha platform, which includes confidential computing. It is designed for large-scale data management and AI model serving.

Discover the synergy between Amun AI and AI4Life, and learn how their collaboration is advancing bioimaging and AI research.

 

Watch the full interview:

Categories
News

New Governance page on the AI4Life website

New Governance page on the AI4Life website

by Beatriz Serrano-Solano

We are pleased to announce a new section on the AI4Life website dedicated to our governance structure. If you want to know who is behind each project body, visit the “Governance” page under the “About Us” section.

On this page, you will find detailed information about:

– Project Management Team: The core team responsible for the day-to-day operations of AI4Life.
– Steering Committee: The advisory group guiding the project’s strategic direction.
– General Assembly: The broader assembly of stakeholders involved in the project.

Additionally, we have two bodies that include external members:

– Executive Board: A team of senior experts providing high-level oversight.
– Open Calls Selection Committee: The committee responsible for evaluating submissions and selecting projects for support.

 

This new governance page aims to provide transparency and insight into the dedicated individuals and committees driving AI4Life forward.

🔗 Visit the Governance Page at https://ai4life.eurobioimaging.eu/governance

Categories
News

Open Call & Challenges Mini-Hackathon in Milan

Open Call & Challenges Mini-Hackathon in Milan

by Wei Ouyang, Estibaliz Gómez de Mariscal, Craig Russell, Vera Galinova, Mehdi Seifi & Beatriz Serrano-Solano

The AI4Life May Hackathon held in Milan ended with successful contributions and progress made across several key areas. Here’s a summary of the event’s highlights:

Consultation Phase of the Second Open Call 

The consultation phase of the second Open Call started during the week of the hackathon. In-person participants engaged in calls with the Open Call applicants to advise them on their projects and better assess the complexity of possible solutions. These consultations addressed a variety of bioimage analysis and deep learning challenges, providing valuable insights and advice to projects from all over Europe. 

Public Challenges

Participants provided critical feedback on the AI4Life-MDC24 challenge pages hosted on the Grand Challenge platform. This included testing data downloads and attempting to train their own denoising models.

Infrastructure Improvement for Better Scalability

A significant milestone was achieved in upgrading Hypha and the BioEngine. This involved:

  • Transitioning the deployment base image from Hypha to the BioEngine.
  • Building a separate Hypha image with necessary plugins to better support the BioEngine.
  • Setting up environment requirements to facilitate this transition, ensuring the new BioEngine image can include additional apps like Cellpose and ImageJ.

Enhancements also included integrating a Ray cluster to manage compute tasks, providing a scalable and efficient alternative to the Kubernetes API. This upgrade is essential for the fine-tuning of models, ensuring better performance and resource management.

Model Uploader

Efforts were concentrated on developing the new model uploader and the model review feature. Detailed coordination plans were discussed for migrating the old collection to the new system, ensuring a smooth transition and improved functionality.

Crowdsourcing Interactive Annotations

Participants worked on the creation of a crowdsourcing annotation tool, enabling collaborative image annotations by multiple users. This tool is accessible here.

Interactive Annotation Tool

Another project focused on an interactive annotation tool based on Kaibu and micro-SAM, resulting in a browser-compatible prototype.

BioImage.IO-Google Colab Project

A project was prototyped to connect the BioEngine with Kaibu for collaborative annotations. This included documenting the bioimageio_core and bioimageio_spec and updating an example notebook for local or Google Colab use, showcasing annotation with BioImageIO-Colab and fine-tuning a Cellpose model.

Thank you to all the participants for making the best out of this event and providing the scientific community with new improvements and achievements.



Categories
News

First AI4Life challenge launched: Denoising microscopy images

First AI4Life challenge launched: Denoising microscopy images

by Vera Galinova and Beatriz Serrano-Solano

We are happy to announce the launch of the first AI4Life challenge aimed at improving denoising techniques for microscopy images.

Noise introduced during the image acquisition process can degrade their quality and complicate interpretation. But deep learning can help with that!

The challenge focuses on unsupervised denoising to be applied to four datasets featuring two types of noise: structured and unstructured. 

To participate, please visit the dedicated website and the Grand Challenge page on which the challenge is hosted:

https://ai4life-mdc24.grand-challenge.org/ai4life-mdc24/

Categories
News

Outcomes of the Second AI4Life Open Call

Outcomes of the Second AI4Life Open Call

 

by Beatriz Serrano-Solano

AI4Life launched its second Open Call on January 22nd, and applications were accepted until March 8th. We’re very happy to share that we received a total of 51 applications for the second AI4Life Open Call!

The applications span a wide range of fields, including developmental and marine biology, as well as cancer research. But that’s not all—we also received submissions from areas such as plant biology, parasitology, ecology, biophysics, microbiology, immunology, and many more.

Why did scientists apply? What challenges are they facing?

Similar to the first open call, most applicants focused on improving their image analysis workflows. However, unlike the first Open Call, we did not offer consultancy as an option to choose, as it is now an implicit step in the process.

Before finalizing project selections, we will hold consultancy calls with a number of project applicants to offer quick guidance that could help researchers find solutions. Projects requiring deep learning support will then be prioritized for the final selection.

How have applicants addressed their analysis problem so far?

Nearly three-quarters of the applicants had already analysed the data, a higher percentage compared to last year. However, only one-fifth of them had not done so yet. The remaining applicants had already analyzed the data and found the outcome satisfactory. Last year, twice the fraction of projects were satisfied with the outcome compared to this time.

When asked about the tools used to analyze their image data, Fiji and ImageJ remain the most popular choices, followed by custom scripts and commercial software.

What kind of data and what format do applicants deal with?

In terms of data format, we observe a change in the trend compared to the first call. Now, 3D images are slightly more frequent than 2D images, reversing the trend we observed last time.

TIFF remains the most popular format, while the second group comprises more proprietary file formats. We are thrilled to observe that the Next Generation File Format (NGFF) has appeared this time!

How much project-relevant data do the applicants bring?

The majority of projects (45 out of 51) manage data in the order of gigabytes or above. Only 6 projects involve data sizes up to a few hundred megabytes. This highlights the prevalence of larger-scale datasets within the applicants, suggesting a growing demand for proper data management and processing capabilities.

Is ground truth readily available?

When asked about the availability and percentage of labelled data, approximately half of the applicants (23 out of 51) reported a lack of sufficient labelled data. Despite a variance in the definition of labelled data compared to the first open call, we observe a comparable trend. Additionally, the proportion of projects indicating high-quality labels has decreased compared to the previous call.

In the application form for this second open call, we introduced an additional question regarding the percentage of labelled data. Over three-quarters of the projects have less than 25% of their data labelled, while approximately 20% of the projects have more than half of their data labelled.

Is the available data openly shareable?

Interestingly, the ratio of projects that can provide access to all their data remains consistent compared to the first open call. This time, we’ve introduced the option of exclusively sharing the controls. The reason behind this decision is that projects unable to share any data are not eligible for support in applying deep learning to their research project.

What’s next?

We have completed the eligibility check and expert reviews for the submitted projects. Projects were reviewed by 17 experts, resulting in an average of 3 reviews per project. The reviews have been aggregated, and scores have been computed to rank the projects based on these reviews. As a result, a preselection of projects has been made, and these applicants will soon be notified.

 

What to expect from the consultation phase?

During the consultation call, selected applicants will have the opportunity to engage with experts who will provide insights, tips, existing tools and recommendations to guide their project. Following the consultation phase, a subset of projects will be chosen to receive expert support based on their potential and need for deep learning support.

Categories
News

Hackathon Summary: BioImage Model Zoo Enhancements

Hackathon Summary: BioImage Model Zoo Enhancements

by Beatriz Serrano-Solano and Joran Deschamps

AI4Life recently organised a hackathon aimed at enhancing the capabilities of the BioImage Model Zoo. Participants from across Europe gathered for a week-long event hosted at EMBL Heidelberg in Germany.

The event kicked off with a round of introductions, allowing participants to outline their personal goals for the week and to form teams that would tackle various project ideas.

Model uploader

One of the key focus areas was to put the final touches to a new model uploader, aimed at simplifying the process of uploading models to the BioImage Model Zoo. The uploader will no longer rely on external platforms like Zenodo; instead, models will be hosted internally and authentication will be requested for contributors who want to upload a model. One of the teams worked to simplify authentication procedures and optimize model uploads to S3 by integrating Google authentication, providing a unified system that would enhance the overall user experience.

Infrastructure improvement

Teams dedicated their efforts to refining Continuous Integration (CI) processes, which have now been migrated to the collection-bioimage-io GitHub repository. The uploader now triggers the CI workflow, automating the process of pushing models to the designated storage location on S3.

JupyterHub and DL4MicEverywhere

Another focus area involved transitioning the infrastructure for JupyterHub from Google Colab to Google Cloud, providing users with a more robust and flexible environment.

Model quantization
Model quantization allows making networks smaller and faster without loss of precision. We held discussions describing current state of the art.. As an example of the performance gains, a 3D Unet model from the BioImage Model Zoo reduced the inference of a batch of images from 60 ms to 30 ms.

Hypha launcher

Hypha can now launch BioEngine (triton-server) on a Slurm cluster using Apptainer. Additionally, a service-id option is now implemented in the BioEngine web client to easily switch the execution backend to high-performance computing (HPC) environments. Furthermore, BioEngine can now be launched on desktop environments.

Model export to new specifications

This team focused on exporting models using the new specifications. Additionally, the team explored approaches to export CellPose models.

Documentation enhancement

This project was split into two phases: firstly, restructuring the current documentation and secondly, creating new needed documentation for the BioImage Model Zoo. Community input and feedback are highly encouraged in this project!

Second Open Call

The deadline for the second Open Call was March 8th. During the hackathon, we had the opportunity to engage with all the reviewers, many of whom were participating in the event. Projects were assigned to each reviewer, kickstarting officially the review process.

Thank you to everyone who contributed either onsite or online. It was a pleasure to work with this engaged group of people. And thank you to the AI4Health innovation cluster for supporting this event. We look forward to meeting you at the next event!