From neutrino physics to cancer detection

It’s not uncommon for developments in particle physics to make their way into medicine. Positron emission tomography (PET), cyclotrons used to generate diagnostic isotopes and hadron therapy, are well-known examples of revolutionary technologies that originated in our quest to understand how matter works. And this transfer of technology is not limited to just hardware: the latest breakthrough technology, Artificial Intelligence (AI), actively developed in particle physics, is also finding applications in other fields.

This time it was thanks to the work of our team of neutrino physicists (past and present), from University of Cambridge, ETH Zurich and University of Cincinnati. We decided to apply a new type of neural network from the reconstruction of neutrino events in liquid argon time-project chambers (TPCs) – used to detect neutrinos at experiments including MicroBooNE and DUNE – to finding tumours in radiological imaging.

Submanifold sparse convolutional neural networks (SSCNs) are a type of image processing algorithm that has found popularity in neutrino experiments due to the typical sparsity of data, where typically only ~1% of readout channels record a signal from an interaction. Standard convolutional neural networks (CNNs), whilst very powerful to identify whether a photograph contains, for instance, a cat or a dog, are inefficient when applied to such sparse images as they perform operations on regions that are all zero-valued (‘empty’). Instead, SSCNs use significantly less memory and take much less processing time for at least the same level of accuracy.

And these advantages can be particularly helpful to enable the deployment of AI tools in clinical environments with limited computational resources, speeding up image analysis while contributing to lower energy consumption during both training and deployment. In this study, we applied SSCNs to computed tomography (CT) imaging, one of the most common image modalities used worldwide for cancer diagnosis and treatment monitoring. Although CT images are not naturally sparse, this imaging modality is a suitable use case for SSCNs. This is possible since the intensities stored in voxels in CT are described by the scale of Hounsfield Units (H.U.) that represents radiodensity, such that different values are indicative of different tissue types.

Example scan showing very good agreement between the high-resolution prediction (bottom) and low-resolution prediction (centre), and the expert radiologist segmentation (top) of the kidneys, tumours and cysts.

Image: Example scan showing very good agreement between the high-resolution prediction (bottom) and low-resolution prediction (centre), and the expert radiologist segmentation (top) of the kidneys, tumours and cysts.

The first step of our method was ‘voxel sparsification’, followed by a two-stage 3D segmentation method in which SSCNs were deployed. In the first stage, a low-resolution sparse network was used to identify a region of interest, a process that in principle can be used in a standalone way for lesion detection. In the second stage, a high-resolution sparse network performed refined semantic segmentation within the region of interest, providing the very precise delineation of organs and tumours needed for further downstream measurements and analysis, which is key to enabling precision medicine and personalised treatments.

As a proof-of-concept, we applied this novel method to the segmentation of kidneys and renal tumours, using only publicly available datasets from the KiTS (Kidney Tumour Segmentation) challenge. The performance we obtained, measured in terms of ‘Dice similarity coefficient’, a metric of overlap between prediction and ground truth widely used in medical segmentation tasks, is competitive with the approaches in the top leaderboard of the latest KiTS challenge, achieving 95.8% for kidney + masses, 85.7% for tumours + cysts, and 80.3% for tumours alone. Crucially, compared to an equivalent dense implementation of the same architecture (UNet), our proposed sparse approach achieves up to a 60% reduction in inference time and up to a 75% reduction in VRAM (video random-access memory) usage across both the CPU and GPU configurations tested.

This study, now published open access in Nature Scientific Reports, provides strong motivation for the application of our SSCN technique to other use cases, and to continue building bridges between particle physics and clinical medicine.

Reference: Alonso-Monsalve, S., Whitehead, L.H., Aurisano, A. and L. Escudero Sanchez, ‘Submanifold sparse convolutional networks for automated 3D segmentation of kidneys and kidney tumours in computed tomography.’ Sci Rep (2026). DOI:10.1038/s41598-026-51801-7

Lorena Escudero Sánchez is a Principal Research Data Scientist at RAND Europe, an affiliated lecturer at the Department of Physics and a visiting researcher at the Department of Radiology.

Leigh Whitehead is a research professor in the High Energy Physics Group working on the DUNE and ProtoDUNE experiments.