Brain MRI can leave radiologists facing difficult distinctions between gliomas and brain metastases, especially when malignant lesions share ring-like enhancement, central necrosis and surrounding oedema. A recent article in npj Digital Medicine presents a deep learning system for detecting brain tumours and distinguishing gliomas from brain metastases on MRI. The evaluation included 3,909 participants from seven centres and used contrast-enhanced T1-weighted imaging with T2-FLAIR sequences. The system combines BTSC-Net, a multi-task model for tumour segmentation and classification, with BTSC-CAD, a computer-aided diagnosis tool that visualises tumour masks and supports radiologists during image review. Histopathological diagnosis remains the standard, but it requires an invasive procedure.
A Unified Model for Detection and Diagnosis
BTSC-Net brings tumour detection and diagnostic classification into a single deep learning framework. The model uses MRI inputs to identify abnormal tumour regions and classify cases as glioma or brain metastasis. Rather than treating segmentation and classification as separate steps, the framework links the two tasks so that shared image features can support both outputs.
The segmentation component combines different image-analysis modules to capture tumour boundaries and regional features. The classification component then uses information from segmentation alongside re-encoded imaging features to generate diagnostic outputs. Training uses a combined loss function intended to support both tumour delineation and classification.
The computer-aided diagnosis system, BTSC-CAD, applies the model through a client-server structure. Its functions include image storage and management, AI-assisted detection and diagnosis, 3D segmentation visualisation, formatted report generation and printing. The interface is designed to show tumour masks, giving radiologists a visual aid during case interpretation. The system was also integrated into a hospital PACS environment, linking the model more closely with routine imaging workflows.
Strong Results Across Internal and External Testing
The model was trained and tested using data collected between January 2015 and January 2024. Cases from three centres formed the training and internal testing datasets, while four further centres provided an external test set. Each case included both MRI sequences. Experienced radiologists manually marked tumour regions, and a senior radiologist reviewed the annotations until consensus was reached. Diagnostic labels were determined with reference to pathology reports.
Must Read: DL Improves Motion-Robust Paediatric Brain MRI
BTSC-Net showed strong tumour detection performance in both test sets. Detection accuracy remained high internally and externally, with only a small decrease between the two groups. The model also performed consistently across age and sex groups. Imaging sequence, magnetic field strength and T2-FLAIR acquisition matrix had limited impact on performance, while thinner contrast-enhanced T1-weighted slices produced the strongest results.
For distinguishing gliomas from brain metastases, BTSC-Net achieved similar diagnostic performance in internal and external testing. The model performed better for gliomas with larger abnormal regions. Compared with several other segmentation and classification approaches, BTSC-Net improved diagnostic accuracy and discriminative performance. Its performance still remained affected by image quality factors such as slice thickness and in-plane resolution.
Support for Junior Radiologists
BTSC-CAD underwent clinical evaluation using the external test set. One senior radiologist and four junior radiologists reviewed cases in a two-phase crossover design. Junior radiologists completed independent reads and assisted reads, with a washout period between phases to reduce recall bias. The comparison assessed detection, diagnosis and reading time.
For tumour detection, BTSC-CAD performed close to the senior radiologist. Junior radiologists had lower baseline detection performance, but their sensitivity and overall diagnostic measures improved with BTSC-CAD assistance. The system helped identify small lesions that were otherwise difficult to detect.
For differentiating gliomas from brain metastases, junior radiologists also showed clear gains with BTSC-CAD assistance. The largest improvements occurred among radiologists with lower baseline performance. Reading time also fell across junior readers, with an average reduction of about one minute. The system did not replace expert interpretation. Diagnostic specificity was slightly lower in some cases, raising concerns about possible overdiagnosis and reinforcing the need for radiologist oversight.
BTSC-Net combines segmentation and classification to support MRI-based detection and diagnosis of gliomas and brain metastases. BTSC-CAD translated the model into a computer-aided diagnosis system with visualisation and reporting functions. Performance remained strong across internal and external testing, and junior radiologists improved with system assistance. The tool still requires manual review before clinical use and applies only to gliomas and brain metastases. Prospective validation under real-world prevalence and open-set conditions remains necessary before broader clinical deployment.
Source: npj Digital Medicine
Image Credit: iStock
References:
Liu X, Lv K, Du P et al. (2026) Multi-task deep learning assists detection and diagnosis of gliomas and brain metastases. npj Digit Med. https://doi.org/10.1038/s41746-026-02759-3
