Estimation of Cardiac Short Axis Slice Levels with a Cascaded Deep Convolutional and Recurrent Neural Network Model

Deep learning for classification of late gadolinium enhancement lesions based on the 16-segment left ventricular model

PURPOSE

This study aimed to develop and validate a deep learning-based method that allows for segmental analysis of myocardial late gadolinium enhancement (LGE) lesions.

METHODS

Cardiac LGE data from 170 patients with coronary artery disease and non-ischemic heart disease were used for training, validation, and testing. Short-axis images were transformed to polar space after identification of the left ventricular (LV) center point and anterior right ventricular (RV) insertion point. Images were obtained after dividing the polar transformed images into segments based on the 16-segment LV model. Five different deep convolutional neural network (CNN) models were developed and validated using the labeled data, where the image after the division corresponded to a segment, and the lesion labeling was based on the 16-segment LV model. Unseen testing data were used to evaluate the performance of the lesion classification.

RESULTS

Without manual lesion segmentation and annotation, the proposed method showed an area under the curve (AUC) of 0.875, and a precision, recall, and F1-score of 0.723, 0.783, and 0.752, respectively for the lesion class when the pretrained ResNet50 model was tested for all slice images. The two pretrained models of ResNet50 and EfficientNet-B0 outperformed the three non-pretrained CNN models in terms of AUCs (0.873-0.875 vs. 0.834-0.841).

CONCLUSION

The proposed method is based on learning a deep CNN model from polar transformed images to predict LGE lesions with good accuracy and does not require time-consuming annotation procedures such as lesion segmentation.

A deep learning model adjusting for infant gender, age, height, and weight to determine whether the individual infant suit ultrasound examination of developmental dysplasia of the hip (DDH)

Objective

To examine the correlation between specific indicators and the quality of hip joint ultrasound images in infants and determine whether the individual infant suit ultrasound examination for developmental dysplasia of the hip (DDH).

Method

We retrospectively selected infants aged 0-6 months, undergone ultrasound imaging of the left hip joint between September 2021 and March 2022 at Shenzhen Children's Hospital. Using the entropy weighting method, weights were assigned to anatomical structures. Moreover, prospective data was collected from infants aged 5-11 months. The left hip joint was imaged, scored and weighted as before. The correlation between the weighted image quality scores and individual indicators were studied, with the last weighted image quality score used as the dependent variable and the individual indicators used as independent variables. A Long-short term memory (LSTM) model was used to fit the data and evaluate its effectiveness. Finally, The randomly selected images were manually measured and compared to measurements made using artificial intelligence (AI).

Results

According to the entropy weight method, the weights of each anatomical structure as follows: bony rim point 0.29, lower iliac limb point 0.41, and glenoid labrum 0.30. The final weighted score for ultrasound image quality is calculated by multiplying each score by its respective weight. Infant gender, age, height, and weight were found to be significantly correlated with the final weighted score of image quality (P < 0.05). The LSTM fitting model had a coefficient of determination (R2) of 0.95. The intra-class correlation coefficient (ICC) for the α and β angles between manual measurement and AI measurement was 0.98 and 0.93, respectively.

Conclusion

The quality of ultrasound images for infants can be influenced by the individual indicators (gender, age, height, and weight). The LSTM model showed good fitting efficiency and can help clinicians select whether the individual infant suit ultrasound examination of DDH.

Evaluation of convolutional neural networks for the detection of inter-breath-hold motion from a stack of cardiac short axis slice images

PURPOSE

This study aimed to develop and validate a deep learning-based method that detects inter-breath-hold motion from an estimated cardiac long axis image reconstructed from a stack of short axis cardiac cine images.

METHODS

Cardiac cine magnetic resonance image data from all short axis slices and 2-/3-/4-chamber long axis slices were considered for the study. Data from 740 subjects were used for model development, and data from 491 subjects were used for testing. The method utilized the slice orientation information to calculate the intersection line of a short axis plane and a long axis plane. An estimated long axis image is shown along with a long axis image as a motion-free reference image, which enables visual assessment of the inter-breath-hold motion from the estimated long axis image. The estimated long axis image was labeled as either a motion-corrupted or a motion-free image. Deep convolutional neural network (CNN) models were developed and validated using the labeled data.

RESULTS

The method was fully automatic in obtaining long axis images reformatted from a 3D stack of short axis slices and predicting the presence/absence of inter-breath-hold motion. The deep CNN model with EfficientNet-B0 as a feature extractor was effective at motion detection with an area under the receiver operating characteristic (AUC) curve of 0.87 for the testing data.

CONCLUSION

The proposed method can automatically assess inter-breath-hold motion in a stack of cardiac cine short axis slices. The method can help prospectively reacquire problematic short axis slices or retrospectively correct motion.