# Dual deep learning approach for non-invasive renal tumour subtyping with VERDICT-MRI

## TL;DR

This study uses a new MRI technique called VERDICT-MRI and deep learning to better distinguish types of kidney tumors non-invasively.

## Contribution

A dual deep learning approach with VERDICT-MRI is introduced for improved renal tumor subtyping and reduced scan time.

## Key findings

- VERDICT-MRI more accurately captures tumor microstructure compared to simpler dMRI models.
- Intracellular and vascular volume fractions differ significantly between cancerous and normal tissue.
- A reduced b-value protocol cuts scan time by over 30 minutes without losing accuracy.

## Abstract

Renal cell carcinomas (RCCs) have multiple subtypes that are difficult to distinguish using imaging alone. This study characterises renal tumour microstructure using diffusion MRI (dMRI) and the Vascular, Extracellular and Restricted Diffusion for Cytometry in Tumours (VERDICT)-MRI framework. Patients were prospectively recruited from the RIM trial (ClinicalTrials.gov: NCT07173140, 20/11/2024). Fourteen patients with 17 renal tumours (including benign and various RCC subtypes) underwent dMRI using nine b-values (0–2500 s/mm²). A three-compartment VERDICT model was fitted with a self-supervised neural network. Compared to simpler dMRI models, VERDICT more accurately captured the diffusion data in tumour and healthy tissue. VERDICT revealed significant differences in intracellular volume fraction between cancerous and normal tissue, and in vascular volume fraction between vascular and non-vascular regions. A feature selection method identified a reduced 4 b-value protocol (b = [70, 150, 1000, 2000]), cutting scan time by over 30 min, enabling more efficient imaging in larger cohorts.