How artificial intelligence (AI) improves image analysis workflows in life sciences

Image acquisition is not the end of the story—it is the starting point.

Once images are captured, researchers typically move into preprocessing to improve quality, then they perform visualization, segmentation, and quantitative analysis to extract meaningful measurements. The goal is simple: turn images into understanding.

In practice, however, image analysis is rarely straightforward. Microscopy datasets can suffer from low signal intensity and poor contrast, and biological structures often can be complex, subtle, and difficult to detect reliably with traditional tools.

Using machine learning and deep learning, AI-based image analysis is changing the workflow by increasing accuracy, reducing manual effort, and extracting more reliable quantitative insights from complex biological data.

How does AI improve image denoising in microscopy?

Noise is a major barrier to accurate image interpretation. Over the past decade, AI-based denoising methods have emerged as powerful approaches for removing unwanted noise while preserving the structural integrity of filaments, organelles, cells, and tissues.

In contrast to classical denoising techniques, which rely on predefined noise models, AI-based methods learn directly from data. This enables them to effectively function as experts trained on different datasets, often delivering more reliable results.

Two commonly used approaches include:

• Unsupervised denoising (Noise2Void) removes noise from images without requiring clean reference images, and it uses information from neighboring pixels.
• Supervised denoising (Noise2Clean) learns to remove noise from images using paired noisy and clean reference images.

Thermo Scientific Amira Software includes multiple AI-based denoising approaches, including pretrained models, Noise to Void, and Noise to Clean. These tools help reveal structures that may be hidden, especially in low-signal datasets.

How does AI improve image segmentation?

Through the use of trained models, AI learns to recognize cellular features, biological heterogeneity, and contrast variations directly from annotated datasets. As a result, segmentation becomes more robust, reproducible, and scalable for large imaging datasets. Compared to traditional methods, this approach represents a major advancement, increasing accuracy while reducing processing time.

Depending on how spatial information is incorporated, AI-based segmentation frameworks can be broadly categorized into three strategies:

• 2D segmentation—processes one slice at a time without considering neighboring slices
• 2.5D segmentation—incorporates information from adjacent slices to improve spatial context
• 3D segmentation—operates on volumetric data and uses full spatial information

As the approach shifts from 2D to 3D, segmentation accuracy generally increases; however, this comes at the cost of requiring larger annotated datasets, longer training times, and greater computational resources.

For this reason, 2.5D segmentation is often considered a practical compromise, offering strong performance while maintaining moderate computational requirements. Figure 2 compares the 3D reconstruction carried out in 2D and in 2.5D, showing that the 2.5D Unet performs better, effectively reducing holes in segmented membrane structures. Beyond visual inspection, model performance should be monitored and evaluated using quantitative metrics to ensure that segmentation results are accurate.

How AI enables more reliable cell segmentation

The integration of AI into segmentation workflows has further expanded toward a cell-level segmentation, enabling the reliable identification of individual cells. This task becomes particularly challenging in densely packed environments, where cells are in close proximity, making boundary delineation difficult.

To address this limitation, CellPose segmentation was developed, providing robust instance segmentation of individual cells even in crowded conditions.

What has emerged is that a significant transformation has occurred in the field of image processing. This shift is not limited to advances in segmentation itself but also includes the development of new annotation tools, as well as the introduction of novel modeling approaches (e.g., shallow and deep learning).

In our on-demand webinar, How AI helps with Image Data Analysis in Life Sciences, we show different AI-based segmentation tools available in Amira Software. Watch this webinar to learn more about AI annotation tools, flexible model training (2D / 2.5D / 3D), shallow vs. deep architectures, as well as model management and metrics.

How does AI enable better data interpretation?

The next step after segmentation is data interpretation, which enables the extraction of meaningful conclusions from the analysis. When segmentation produces a large number of objects, strategies for organization and classification become essential to support systematic study. In this context, artificial intelligence has enabled the development of advanced methods capable of classifying segmented elements with high accuracy and efficiency.

Explore the AI object classification implemented in Amira Software as shown in Figure 4.

Conclusion

Less manual work. More reliable results.

The implementation of artificial intelligence tools within the image analysis workflow has enabled the optimization of image processing steps, resulting in improved accuracy.

Amira Software positions these AI capabilities as practical, guided tools with no coding required, and often effective with only a small number of annotated slices to start.

Watch the full on-demand webinar to learn more about AI-powered image analysis workflows in Thermo Scientific Amira Software.

Interested in discussing your image analysis challenges? Our experts are here to help whenever you’re ready.