732 
A deep learning-based diagnostic tool can accurately distinguish between three of the most common brain tumours, outperforming conventional techniques, Spanish researchers report.
Glioblastoma multiforme, brain metastasis from solid tumours and primary central nervous system lymphoma accounted for up to 70% of all brain malignancies, the researchers wrote in the journal Cell Reports Medicine.
Each type of tumour required a distinct therapeutic approach, but on imaging they appeared similar, making it difficult to distinguish each type.
Corresponding study author Dr Raquel Perez-Lopez, head of Vall d’Hebron Insitute of Oncology’s (VHIO) Radiomics Group, said magnetic resonance imaging (MRI) was currently used for non-invasive differential diagnosis.
‘However, a definitive diagnosis often requires neurosurgical interventions that compromise the quality of life of patients,’ she said.
To overcome these challenges, researchers from VHIO Radiomics Group and the Neuro-Radiology Unit at Bellvitge University Hospital, developed a deep learning-based tool, leveraging spatial and temporal information from dynamic susceptibility contrast (DSC) perfusion MRI to assist in classifying brain tumours.
‘In DSC, every voxel in the image yields a unique dynamic curve that describes the temporal evolution of the T2∗-weighted signal intensity and reflects local tissue vascular properties,’ the researchers wrote.
‘The standard approach to analyse DSC is to derive metrics such as the relative cerebral blood volume (rCBV) and the percentage of signal recovery (PSR) which both simplify the dynamic signal.’
The tool, known as Diagnosis in Susceptibility Contrast Enhancing Regions for Neuro-oncology (DISCERN), was trained to recognise characteristics of these common brain tumours using approximately 50,000 voxels from the DSC perfusion MRI images of 40 histology-confirmed patients.
It was then tested on 400 additional cases, plus an external validation cohort of 128 patients, with the researchers reporting it reached a three-way accuracy of 0.78, superior to the conventional MRI metrics of rCBV and PSR.
The tool required approximately two minutes to process a new case, the researchers said, and achieved optimal performance through training with a limited number of scans from somewhere in the order of 30 to 40 cases.
‘These data underscore the potential of DISCERN for differentiating among the three most common clinical diagnostic challenges in patients with enhancing brain lesions,’ they concluded.
Study co-author Dr Albert Pons-Escoda, a clinical neuroradiologist at Bellvitge University Hospital in Barcelona, said the work was the result of more than five years of research focused on identifying innovative magnetic resonance perfusion imaging biomarkers for differential diagnosis of brain tumors.
‘This present study integrates insights generated by other previous research projects on artificial intelligence, resulting in the development of software that automates presurgical diagnostic classification with very good precision, while facilitating its clinical applicability with a user-friendly interface for clinicians,’ he said.
AI models are increasingly being used for a range of healthcare applications. Deep learning has recently enabled scientists to accurately predict four subtypes of Parkinson’s disease based on images of patient-derived stem cells.
In a systemic review and meta-analysis published last year, Canadian researchers found that the performance of an AI model for the diagnosis of hip fractures was comparable with that of expert radiologists and surgeons.
And in research presented at the European Academy of Dermatology and Venereology Congress 2023, showed that the use of artificial intelligence software had a 100% detection rate for melanoma and saved over 1,000 face-to-face secondary care consultations during a 10-month period.
