An international team of researchers, in a comprehensive review, has described the combination of autofluorescence imaging (AF) with deep learning methods as a promising approach for the early, marker-free detection of diseases. The integration makes it possible to visualize metabolic and microstructural changes in tissue long before classic morphological changes occur. The review was published in the journal Frontiers in Artificial Intelligence.
Background
Many diseases, including precancerous conditions, diabetic complications, and chronic inflammation, begin with biochemical and metabolic changes in the tissue. These early changes are often not yet visible with conventional imaging techniques. Autofluorescence imaging uses the natural fluorescence of endogenous molecules (e.g., NADH, FAD, collagen, lipofuscin) to obtain metabolic and structural information without dyes or contrast agents. Deep learning algorithms can analyze these complex optical signals and recognize patterns that remain hidden from the human eye.
Technological Basis
Autofluorescence arises from endogenous fluorophores that provide insights into cellular metabolism, redox status, mitochondrial function, and the composition of the extracellular matrix. Particularly informative are the redox cofactors NADH and FAD, as well as structural proteins like collagen and elastin.
Deep learning models (especially Convolutional Neural Networks and transformer architectures) are used to extract clinically relevant information from hyperspectral, time-resolved (FLIM), or multispectral AF data. The combination of both technologies enables functional, biochemically based imaging instead of purely morphological diagnostics.
Clinical Applications
The AF-DL combination is already being investigated in several medical fields:
- Ophthalmology: Fundus autofluorescence (FAF) in combination with DL for the classification of inherited retinal diseases and for predicting the progression of geographic atrophy in age-related macular degeneration.
- Otorhinolaryngology and Thyroid Surgery: Near-infrared autofluorescence for intraoperative identification of parathyroid glands with high accuracy.
- Lungs: Autofluorescence bronchoscopy for improved detection of early-stage lung cancers and precancerous lesions.
- Digestive Tract: Detection of Barrett's esophagus, colorectal polyps, and inflammatory bowel diseases.
- Urology: Non-invasive assessment of kidney diseases via autofluorescent cells in urine and improved bladder cancer detection.
- Dermatology: Distinguishing melanomas and moles, and measuring advanced glycation end products (AGEs) as markers of metabolic stress.
In several studies, the combination of AF and deep learning significantly improved diagnostic accuracy compared to pure autofluorescence or white light imaging.
Challenges
Despite the promising results, significant hurdles remain:
- Strong dependence of signals on device type, illumination, and tissue optics
- Heterogeneity of datasets and lack of standardization
- Limited generalizability of models to other centers
- High annotation effort and interpretability of AI models
- Regulatory and workflow-related issues for real-time applications in the operating room or during endoscopy
Outlook
The authors see the future in the development of multidimensional, hardware-independent systems that combine spectral, temporal, and spatial information. Important next steps include standardized acquisition protocols, prospective multicenter studies, and seamless integration into clinical workflows. In the long term, the AF-DL combination could lead to a new class of functional imaging biomarkers that enable earlier, more precise, and less invasive diagnostics.
FAQ
What is Autofluorescence Imaging?
A marker-free method that utilizes the natural fluorescence of the body's own molecules (e.g., NADH, FAD, collagen) to visualize metabolic and structural tissue changes.
Why is Deep Learning Needed?
Autofluorescence signals are complex and overlapping. AI models can detect subtle patterns that are difficult or impossible to interpret with conventional analysis methods.
In which areas is the technique already being used?
Applications in ophthalmology, thyroid and parathyroid surgery, and for the early endoscopic detection of tumors in the head and neck area and the intestines are particularly advanced.
What advantages does the combination offer?
Earlier detection of changes, better differentiation of benign and malignant lesions, assistance with intraoperative navigation, and potentially a reduction in unnecessary biopsies.
When can widespread clinical use be expected?
Some applications (e.g., parathyroid identification) are already in use in specialized centers. However, larger prospective studies and standardization are still necessary for many other indications.
