The power of the radiologist’s last word: can deep learning models accurately differentiate between high-grade gliomas and metastasis through natural language processing on radiology reports?

Excerpt

The accurate diagnosis of high-grade gliomas (HGGs) and brain metastasis is important due to its profound impact on patient outcomes and management. Distinguishing between these two conditions plays a crucial role in prognostic evaluation and therapeutic planning, with HGGs often requiring neurosurgical resection as the primary approach. Achieving the highest diagnostic accuracy through non-invasive techniques remains a top priority in clinical practice. While histopathology remains the gold standard, the presurgical approach heavily relies on neuroimaging, with magnetic resonance imaging (MRI) as a leading tool. However, the heterogeneity of both tumor types poses challenges, and conventional MRI alone may not always provide a reliable diagnosis. To address this limitation, advanced MRI techniques, such as diffusion tensor imaging (DTI), dynamic contrast-enhanced (DCE)-MRI, MR spectroscopy, and amide proton transfer (APT)-MRI, have emerged as valuable supplements to morphological MRI, enabling improved diagnosis [ 1]. These techniques go beyond morphology to offer information on the metabolic, cellular, or vascular environment levels, providing additional information for accurate classification. However, it is essential to acknowledge that implementing these advanced imaging techniques can be challenging. They typically require higher magnetic field MRI scanners, specialized sequences, improved imaging qualities, and sophisticated post-processing steps, making them inaccessible for many non-academic medical institutes. Consequently, the widespread clinical translation of these advanced imaging techniques is currently hindered, limiting their routine use in clinical practice. …