How to train your AI microscope

Image: Drosophila embryo imaged on a SIMView lightsheet microscope with a 16x 0.8NA water objective every 30 seconds. Cells are tracked using Aivia’s 3D Object Tracking recipe. Credit: Philipp Keller, HHMI Janelia Farms Research Campus, Ashburn VA; and the Cell Tracking Challenge.

The future of microscopy lies in the insights that technology solutions can provide to scientists and researchers. The benefits this will bring to the industry as a whole will be truly transformational.

Artificial Intelligence (AI) is increasingly being deployed across the spectrum of industry sectors to transform outdated workflows, eliminate human error, supercharge productivity, and unlock remarkable new opportunities. Yet despite the revolution that is underway, AI continues to suffer from hype cycles, misunderstanding and misrepresentation. The uncertainty this creates has hindered its adoption – especially in more traditional sectors.

Life sciences is one of those industries. There is safety and comfort in the ‘tried and tested’ – if it’s been used thousands of times before, you know it’s not going to undermine the validity of your research. As a result, the life sciences research market has been slow to adopt new technologies and approaches. But there is growing demand for increased speed to market, and R&D has a key role to play in this. The Covid-19 pandemic has shown us that we can break records for bringing a new treatment with the right funding, collaboration, and digital innovation. This needs to be the catalyst for a radical rethink about the role that AI can play in our industry and the benefits it could bring to life sciences and biopharma companies alike.

Image: Two-colour organoid cluster, stained with DAPI – Nucleus, GFP – Plasma Membrane. Sample courtesy of Dana Krauß, Cancer Research Institute, Medical University Vienna, Austria.

Microscopy is one area of life sciences where we are seeing real innovation in this space. We are in the middle of a period of disruptive transformation where the role of the technology is changing irreversibly. The microscope used to be there to create the best image possible and the focus of innovation and investment was on incremental improvement of the hardware to achieve this. Image quality and resolution are of course still vital, but it’s the insight – what the technology can tell you about the image or images – that provides real value and holds so much potential.

Image analysis has often been an inefficient and tedious part of the job for any scientist who works with microscopes. To understand how a cancerous cell metastasizes, for example, could require hundreds or thousands of images to be reviewed one after the other. For a human, the task is extremely time-consuming, adding to the already huge administrative burden with which all researchers currently struggle. It is also littered with opportunities for mistakes – either through human error, or simply because it is impossible for humans to process this quantity of data effectively. To stay ahead of this trend, over the past five years we acquired and developed a software solution to enable scientists to create AI models for autonomous image analysis. Basic information must be input, and initial analysis undertaken to train the model, but once this has been done the system begins to learn for itself – creating a predictive model, then analysing thousands of images and showing the user the results.

Image: Single timepoint of a time-lapse recording of mammary epithelial microspheroid cultured in 3D highlighting individual mitotic events. Data courtesy of the intelligent imaging group – B. Eismann and C. Conrad, BioQuant / DKFZ Heidelberg, Germany.

Properly trained AI systems can become more efficient than humans at recognising patterns in complex images and correlating them to specific disease states (for example, at analysing fundus photographs to detect diabetic eye disease). But humans are still better at understanding and explaining what these findings actually mean.   

Image: Thy1-EGFP labelled neurons in whole mouse brain processed using the PEGASOS 2 tissue clearing method, imaged on a Leica SP8 confocal microscope, 40x /1.3NA objective. Credit: Hu Zhao, Chinese Institute for Brain Research, Beijing, China.

The future of microscopy lies in the insights that technology solutions can provide to scientists and researchers. The benefits this will bring to the industry as a whole will be truly transformational. Scientists can eliminate most of the burden associated with image analysis, freeing them up to focus on creative tasks that require the intelligent integration of complex information from multiple domains (e.g., image, spatial, genomic, and proteomic data). This is something AI is still not good at and is unlikely to catch up to humans any time soon. If we achieve this shift from tedious admin to critical thinking, we will be able to do better science and do so faster, and this will translate into better quality treatments, brought to market faster.

The technology is already there and improving all the time. We now need to focus on providing the education and training scientists need to understand the capabilities and limitations of AI-powered tools and how they are best applied, so they can confidently communicate how AI was used to help reach their conclusions. It will take time, but make no mistake in five years this will be the norm and we will be wondering what took us so long, when the benefits to the industry, to customers, and to patients are so clear.  

Author: Luciano Lucas is Director, Leica Aivia at Leica Microsystems leica-microsystems.com