DINOv2 framework predicts cardiotoxicity from echocardiograms with 70% AUC

A new paper from Xiaotong Zhang and colleagues introduces a unified framework built on DINOv2 to tackle three complementary cardio-oncology tasks: estimating left ventricular ejection fraction (LVEF), classifying global longitudinal strain (GLS) dysfunction, and predicting early cardiotoxicity. DINOv2 is a self-supervised vision transformer that provides strong general image representations. The team froze the backbone and added parameter-efficient Low-Rank Adaptation (LoRA) alongside temporal aggregation to learn task-specialized features without fine-tuning the entire network.

The framework was evaluated on a patient-level split of 1,203 training videos from 237 patients and 300 validation videos from 59 independent patients. It achieved a mean absolute error of 5.03% for LVEF estimation, an AUC-ROC of 76.48% for GLS dysfunction classification, and an AUC-ROC of 70.26% for cardiotoxicity prediction. Notably, the inference pipeline operates without cardiac cycle segmentation or explicit end-diastolic/end-systolic annotations, simplifying real-world deployment. For Task 1, a specialized ED/ES-guided 2D/3D hybrid multi-view regression model further lowered the MAE to 4.64%.