AI revolution in medical imaging: Tracking biological changes in real-time

Medical imaging has long been a crucial tool for monitoring biological changes over time. From tracking embryo development to assessing wound healing and brain aging, longitudinal imaging helps researchers and clinicians detect critical changes and disease progression. Traditional approaches to analyzing these images rely on manual preprocessing, segmentation, and statistical modeling. However, these methods are often labor-intensive and prone to inconsistencies.

A recent study titled “Learning-based Inference of Longitudinal Image Changes: Applications in Embryo Development, Wound Healing, and Aging Brain” by Heejong Kim, Batuhan K. Karaman, Qingyu Zhao, Alan Q. Wang, and Mert R. Sabuncu, published in PNAS (2025), introduces a new machine-learning framework called LILAC (Learning-based Inference of Longitudinal imAge Changes). This deep-learning model automates the detection of meaningful temporal changes in medical imaging, eliminating the need for extensive manual processing. By using convolutional neural networks (CNNs) in a Siamese architecture, LILAC effectively tracks, quantifies, and localizes significant changes in biological structures over time.

Advancing longitudinal image analysis with AI

Longitudinal imaging provides insights into progressive biological processes, such as tumor growth, neurodegeneration, or wound healing. Traditional analysis methods require custom pipelines to eliminate noise and extract meaningful features, but these approaches rely heavily on user-defined variables and predefined regions of interest. This limits their flexibility and can introduce biases.

LILAC presents an alternative by applying deep learning to directly analyze image pairs, learning the differences between them without predefined assumptions. It operates in three key modes:

1. LILAC-o (ordering mode): Predicts the correct chronological order of two images in a sequence.
2. LILAC-t (time-interval prediction mode): Estimates the time difference between two images.
3. LILAC-s (clinical score prediction mode): Predicts changes in clinical indicators, such as cognitive decline in Alzheimer’s patients.

The model is designed to ignore irrelevant variations while focusing on the most meaningful changes, making it more robust than traditional segmentation-based approaches. The study demonstrated LILAC’s effectiveness across multiple medical applications, from tracking embryo development to analyzing brain aging patterns.

Real-world applications: From embryo development to brain aging

The study applied LILAC to four different medical imaging datasets: embryo development, wound healing, healthy brain aging, and mild cognitive impairment (MCI). Each dataset posed unique challenges, but LILAC consistently outperformed traditional methods in detecting and quantifying temporal changes.

For embryo development, the researchers trained LILAC-o to determine the correct order of images in an embryo time-lapse dataset. The model achieved a 98.9% accuracy rate, successfully identifying subtle changes in cell division phases. The Grad-CAM visualization technique further revealed how LILAC identified key regions of interest during embryonic growth.

In wound healing, LILAC-o analyzed time-lapse images of cell migration, achieving a Pearson correlation coefficient of 0.875 with actual healing time. The model effectively distinguished between treated and untreated wounds, demonstrating its ability to track therapeutic effects.

For brain aging, LILAC-t was used to predict the time intervals between MRI scans of healthy aging subjects. The model achieved an RMSE of 1.825 years, significantly outperforming single-image regression models. It also revealed individualized aging patterns, highlighting the variability in brain structure changes among different individuals.

In mild cognitive impairment (MCI), LILAC-s was used to predict changes in the Dementia Rating Scale Sum of Boxes (CDRSB). The model effectively disentangled age-related changes from neurodegeneration, allowing for more precise tracking of Alzheimer’s progression. The heatmaps generated by LILAC showed different patterns of brain deterioration across patients, emphasizing the heterogeneous nature of neurodegenerative diseases.

Challenges and future directions in AI-driven longitudinal imaging

Despite its strong performance, LILAC still faces challenges that need to be addressed in future research. One limitation is data dependency - while the model effectively learns from longitudinal image pairs, its accuracy depends on the quality and quantity of the training data. In cases where datasets are limited or contain biases, the model’s generalization ability could be impacted.

Another challenge is model interpretability. While LILAC improves on traditional segmentation-based methods, its deep-learning architecture remains a black box to some extent. The study leveraged Grad-CAM visualizations to highlight key regions of interest, but further research is needed to improve AI explainability in medical imaging.

Computational efficiency is also a concern. LILAC relies on CNN-based Siamese architectures, which require significant processing power. Future developments could explore transformer-based models or self-supervised learning techniques to enhance efficiency and scalability.

Future of AI in longitudinal medical imaging

The introduction of LILAC represents a significant leap forward in AI-powered longitudinal image analysis. By eliminating the need for manual preprocessing and handcrafted feature extraction, LILAC makes it easier to track disease progression, assess treatment effectiveness, and study biological development over time.

Moving forward, AI-driven models like LILAC could revolutionize clinical decision-making, enabling physicians to detect early signs of disease and monitor treatment responses more effectively. The study suggests potential applications in cancer monitoring, regenerative medicine, and precision neurology, where detecting small yet meaningful changes is critical.

Ultimately, deep learning is reshaping the way medical imaging is analyzed, providing more accurate, efficient, and scalable solutions for studying temporal biological processes. As AI continues to evolve, tools like LILAC will play a pivotal role in advancing personalized medicine and improving patient outcomes worldwide.