Machine Learning

The Machine Learning team concentrates on the creation of computer-aided diagnostics of diseases, with a special focus on cancer, using fluorescence lifetime endo/microscopic images (FLIM), via image processing, image analysis, and machine learning.

The Proteus team has developed novel optical imaging platforms, and those platforms have been applied to a wide range of biological models, such as lung bacteria/cancer imaging. With advanced post-processing approaches, a large collection of FLIM images have been generated for research, engineering, and clinical purposes. The primary objectives of the machine learning team include, but are not limited to:

Image processing and analysis

Working with the Signal Processing team, we are actively seeking advanced methodologies to continuously improve the quality of the generated images for better human and machine perception. This is particularly needed for in-vivo in-situ clinical applications to deliver real-time capability of FLIM systems.

Machine learning and deep learning

Figure 1: Contrast of averaged lifetime (third column) of normal (first row) and cancerous (second row) lung tissue by histogramming lifetime images (second column) derived from intensity images (first column), along with histology images (fourth column) as the ground truth.

Conventionally, FLIM-based cancer detection relies on lifetime contrast derived by statistical methods with auxiliary information, such as averaged lifetime by histogramming lifetime images, along with the corresponding histological images as the ground truth. Machine learning (ML), particularly deep learning (DL), has revolutionised bio/medical image processing in problems such as classification and segmentation. The team concentrates on investigations of advanced ML/DL approaches, including novel architectures and methodologies, on FLIM-based automatic detection of cancer. We are particularly interested in:

  • ML/DL based cancer classification of different types, including early-stage cancer diagnosis
  • Multi-modality image registration and segmentation: comparing FLIM images with ground truth, eg histological images, is an intrinsic way to correctly interpret FLIM images. Therefore, registration of FLIM images with histological images is of great importance for the interpretation. With the success of the registration, FLIM images can be further segmented to reveal the correlations between fluorescence lifetime and unhealthy tissue at molecular level

Ultimately, ML/DL approaches are expected to help biologists and clinicians to understand underlying reasons of the variety of fluorescence lifetime from the biomedical and clinical points of view, which, in turn, helps improve the performance of ML/DL approaches for reliable diagnostics.

FLIM database for lung cancer binary classification

The biggest barrier to the application of ML, particularly, DL approaches to FLIM-based cancer discrimination is the lack of a publicly available database suitable for the purpose. To eliminate the gap and push forward this special area, we open-source our FLIM database.

So far, over 100,000 FLIM images have been collected on 18 pair of cancerous/normal tissue from 18 patients, using a fibre-based custom FLIM system. We selected over 60,000 images suitable for the classification. Since experiments are still ongoing, the database will be gradually expanded over time.

The database will be available after further cleaning!

The Machine Learning team

Qiang Wang (corresponding contact:

Marta Vallejo (webpage:

James R Hopgood (webpage: