External validation of multitask deep learning for carotid plaque detection: anatomical auxiliary tasks shape performance trade-offs
Related Experiment Videos
Haichao Shi1, Zeyu Lai2, Fengqin Shen3
1Department of Ultrasound Medicine, Zhejiang Provincial People's Hospital, Affiliated People's Hospital, Hangzhou Medical College, Hangzhou, 310014, Zhejiang, China.
Purpose:
Carotid plaque screening using ultrasound is challenged by class imbalance and ambiguous lesion boundaries. However, how different anatomical auxiliary tasks influence carotid plaque detection performance and clinical operating characteristics remains insufficiently investigated. We aimed to externally evaluate multitask deep learning models with different anatomical auxiliary tasks and investigate their effects on clinical performance profiles.
Methods:
We developed a multitask network incorporating three anatomical auxiliary segmentation targets (vessel lumen, intima-media thickness (IMT), and plaque), together with a classification-only baseline. The model was developed on 1297 images from 262 patients and evaluated on an external cohort of 1091 images from 263 patients. To ensure fair comparison, we froze the encoder of each variant and retrained an identical classification head on the internal training set, and then applied a fixed threshold determined via Youden index maximization for each model on the internal validation set to the external cohort.
Results:
In external validation (263 patients, 94 plaque-positive), all multitask models outperformed the classification-only baseline (AUC 0.615, p < 0.0001), while showing comparable AUCs among themselves (lumen-guided 0.852, IMT-guided 0.853, plaque-guided 0.858; all p > 0.76). Despite comparable discrimination, they exhibited distinct clinical profiles: the IMT-guided model achieved the highest sensitivity (SEN 0.830), the lumen-guided model achieved the highest specificity (SPE 0.864), and the plaque-guided model offered more balanced performance (SEN 0.723, SPE 0.822). Radiofrequency balloon ablation significantly reduced external AUC for the lumen-guided (0.852 vs 0.790; p=0.0145) and plaque-guided (0.858 vs 0.796; p=0.0091) models, but not for the IMT-guided model (p=0.2171), indicating task-dependent effects.
Conclusion:
The choice of auxiliary task does not confer universal superiority in discrimination but instead determines the model's clinical orientation, allowing task-specific adaptation to different screening or confirmatory scenarios.