Radiology — Weekly Evidence Brief

Deep learning models are increasingly enhancing diagnostic and prognostic capabilities across various medical specialties. In bladder cancer, an MRI-based multimodal fusion deep learning model (MF-DLM) demonstrated superior performance in predicting overall survival, outperforming traditional pathological T stage and identifying patients most likely to benefit from perioperative therapy 1. Similarly, a multitask deep learning model leveraging routinely acquired CT scans improved cross-sectional and longitudinal vertebral fracture prediction, surpassing both bone-only models and the traditional FRAX tool, offering opportunities for earlier identification and intervention 2. Furthermore, in nasopharyngeal carcinoma, a combined radiomics model achieved high accuracy (AUC 0.90) in predicting local recurrence after intensity-modulated radiation therapy, guiding personalized treatment strategies 3.

Beyond enhanced prediction, AI is also making significant strides in improving patient safety and optimizing diagnostic procedures. A notable development includes deep learning models that successfully restore standard-dose T1ce images from low-dose (10-30%) gadolinium-enhanced MRI of the cerebellopontine angle cistern 4. This capability enables accurate lesion detection and diagnostic characterization with substantially reduced contrast agent, addressing concerns about gadolinium retention and nephrogenic systemic fibrosis. In prostate cancer diagnosis, extending post-MRI PSA density thresholds to ">=0.20 ng/ml2" for men with PI-RADS 3 lesions safely reduced unproductive negative biopsies and detection of indolent GG1 cancers, while maintaining a low rate of undetected significant cancers 5.

AI-driven imaging tools are proving invaluable for personalized management and monitoring of chronic and complex conditions. For patients with relapsing-remitting multiple sclerosis (RRMS), nomograms integrating clinical indicators, cortical morphometric features, and 3D multi-parametric MRI radiomics offer individualized predictions for disability progression and cognitive worsening, guiding early interventions 6. In transthyretin cardiac amyloidosis (ATTR-CM), AI-quantified SPECT/CT markers were identified as potential early indicators of treatment response, with significant reductions in markers observed after disease-modifying therapy, aiding in personalized treatment strategies 7. Additionally, in diabetes, quantitative coronary plaque burden from CTA, rather than myocardial perfusion, emerged as an important predictor of long-term cardiovascular outcomes, indicating that even with normal perfusion, diabetic patients can have higher event risk due to increased plaque 8.

The precision of AI in medical imaging extends to challenging diagnostic scenarios. For intracranial primary tumors, a novel self-refining framework using panoptic segmentation and cross-modality attention improves detection of diagnostically difficult cases and provides anatomically plausible segmentations for surgical planning 9. Similarly, a multi-modal MRI-based deep learning framework (ISMF-Net) significantly enhanced the diagnostic accuracy of intraspinal tumors, particularly for junior radiologists, by integrating various MRI features and clinical data 10. In pediatric cardiology, an automated echocardiogram strain analysis system, Motion-Echo, demonstrated superior performance in detecting cancer therapy-related cardiac dysfunction, myocardial infarction, and left ventricular ejection fraction decline, facilitating earlier detection and improved patient outcomes 11. Furthermore, a deep learning model for renal cell carcinoma showed superior performance over clinical experts in predicting renal sinus invasion from CT images, offering a more precise assessment tool for surgical planning 12.