Back to Projects

Intra-operative probe design and image processing optimisation with deep learning for in-vivo and ex-vivo detection of cancerous tissue

Project ID: 2019_S04

Joint 1st supervisor: Andrew Reader, King’s College London
Joint 1st supervisor: Paul Marsden, King’s College London
2nd supervisor: Kunal Vyas, Lightpoint Medical
Industrial links: Lightpoint Medical

This project involves innovative probe design optimisation, advanced physics modelling and deep-learned image and signal processing to advance the state-of-the-art in intra-operative probes for cancer surgery.

At present, cancer surgery is often unsuccessful, either resulting in the need for multiple subsequent operations or else increasing the need for additional drug treatment or radiotherapy. For example, approximately 25% of patients undergoing surgery for prostate cancer will have a positive surgical margin which is an indicator of incomplete cancer removal. Surgery is unsuccessful so often because surgeons lack a tool to detect cancerous tissue in real time during surgery.

This pressing need can be met through the development of intra-operative technology for detecting radiopharmaceutical tracers which are currently used for pre-operative PET and SPECT scans. These novel intra-operative techniques pose multiple engineering challenges due to the constraints of time, space and limited radioactivity concentrations.  There are also physics modelling challenges due to the need to collect and interpret complex signals arising from radiation absorption, scattering and the presence of interference.

This research project will build both a theoretical and practical understanding of two key and complementary approaches: 1) in-vivo detection of cancer during laparoscopic surgery and 2) ex-vivo detection of cancerous margins on samples. This will involve Monte Carlo simulations of the radiation physics (e.g. GEANT4) and the probe, investigating sources of interference, as well as developing deep-learning assisted image and signal processing techniques (e.g. in MATLAB/Python) in order to exploit all measured data to its fullest informative extent. Specifically, through use of high quality training dataset pairs (low count and high count data), deep learning will allow powerful noise reduction, and even the possibility of enhanced spatial resolution, through use of generative modelling.

The project has scope to extend all the way from probe hardware design optimisation through to development of AI-assisted real-time surgical imaging capabilities.

Promising advances will be further investigated through experiments, prototyping and the use of real data as well as images from medical instruments. This work will be done in close collaboration with industry, with the student also spending time at Lightpoint Medical.

Probe design optimisation will involve experimental and analysis work to understand and evaluate methods and detectors for the detection and identification of internal conversion electrons from 99mTc. This process is attractive as 99mTc is the most widely used radionuclide for nuclear medicine studies, and so complex regulatory issues can be avoided. Reliable detection, identification and localisation these electrons poses many challenges associated with low internal conversion (IC) electron yield, electron energy loss in tissue, and gamma and electron background (in addition to the constraints of laparoscopy).

These issues will be characterised experimentally and with reference to appropriate models, initially to obtain a good understanding of the processes and trade-offs involved and subsequently in the context of workable probe systems. Experiments will be performed initially in the context of an existing Lightpoint prototype device that uses a CMOS sensor to directly image the IC electrons. However other detectors types in various configurations may also be feasible and will be investigated and evaluated.

Back to Projects