Representative Data Selection for Efficient Medical Incremental Learning.

To train a deep neural network relies on a large amount of annotated data. In special scenarios like industry defect detection and medical imaging, it is hard to collect sufficient labeled data all at once. Newly annotated data may arrive incrementally. In practice, we also prefer our target model to improve its capability gradually as new data comes in by quick re-training. This work tackles this problem from a data selection prospective by constraining ourselves to always retrain the target model with a fix amount of data after new data comes in. A variational autoencoder (VAE) and an adversarial network are combined for data selection, achieving fast model retraining. This enables the target model to continually learn from a small training set while not losing the information learned from previous iterations, thus incrementally adapting itself to new-coming data. We validate our framework on the LGG Segmentation dataset for the semantic segmentation task.Clinical relevance- The proposed VAE-based data selection model combined with adversarial training can choose a representative and reliable subset of data for time-efficient medical incremental learning. Users can immediately see the improvement of the medical segmentation model whenever new annotated images are contributed (after a few minutes of model retraining).

Automated Stanford classification of aortic dissection using a 2-step hierarchical neural network at computed tomography angiography.

OBJECTIVES: This study aimed to evaluate the feasibility of automatic Stanford classification of classic aortic dissection (AD) using a 2-step hierarchical neural network. METHODS: Between 2015 and 2019, 130 arterial phase series (57 type A, 43 type B, and 30 negative cases) in aortic CTA were collected for the training and validation. A 2-step hierarchical model was built including the first step detecting AD and the second step predicting the probability (0-1) of Stanford types. The model's performance was evaluated with an off-line prospective test in 2020. The sensitivity and specificity for Stanford type A, type B, and no AD (Sens A, B, N and Spec A, B, N, respectively) and Cohen's kappa were reported. RESULTS: Of 298 cases (22 with type A, 29 with type B, and 247 without AD) in the off-line prospective test, the Sens A, Sens B, and Sens N were 95.45% (95% confidence interval [CI], 77.16-99.88%), 79.31% (95% CI, 60.28-92.01%), and 93.52% (95% CI, 89.69-96.25%), respectively. The Spec A, Spec B, and Spec N were 98.55% (95% CI, 96.33-99.60%), 94.05% (95% CI, 90.52-96.56%), and 94.12% (95% CI, 83.76-98.77%), respectively. The classification rate achieved 92.28% (95% CI, 88.64-95.04%). The Cohen's kappa was 0.766 (95% CI, 0.68-0.85; p < 0.001). CONCLUSIONS: Stanford classification of classic AD can be determined by a 2-step hierarchical neural network with high sensitivity and specificity of type A and high specificity in type B and no AD. KEY POINTS: • The Stanford classification for aortic dissection is widely adopted and divides it into Stanford type A and type B based on the ascending thoracic aorta dissected or not. • The 2-step hierarchical neural network for Stanford classification of classic aortic dissection achieved high sensitivity (95.45%) and specificity (98.55%) of type A and high specificity in type B and no aortic dissection (94.05% and 94.12%, respectively) in 298 test cases. • The 2-step hierarchical neural network demonstrated moderate agreement (Cohen's kappa: 0.766, p < 0.001) with cardiovascular radiologists in detection and Stanford classification of classic aortic dissection in 298 test cases.