Artificial Intelligence (AI) technologies have recently been applied to medical imaging for diagnostic support. With respect to fetal ultrasound screening of congenital heart disease (CHD), it is still challenging to achieve consistently accurate diagnoses owing to its manual operation and the technical differences among examiners. Hence, we proposed an architecture of Supervised Object detection with Normal data Only (SONO), based on a convolutional neural network (CNN), to detect cardiac substructures and structural abnormalities in fetal ultrasound videos. We used a barcode-like timeline to visualize the probability of detection and calculated an abnormality score of each video. Performance evaluations of detecting cardiac structural abnormalities utilized videos of sequential cross-sections around a four-chamber view (Heart) and three-vessel trachea view (Vessels). The mean value of abnormality scores in CHD cases was significantly higher than normal cases (p < 0.001). The areas under the receiver operating characteristic curve in Heart and Vessels produced by SONO were 0.787 and 0.891, respectively, higher than the other conventional algorithms. SONO achieves an automatic detection of each cardiac substructure in fetal ultrasound videos, and shows an applicability to detect cardiac structural abnormalities. The barcode-like timeline is informative for examiners to capture the clinical characteristic of each case, and it is also expected to acquire one of the important features in the field of medical AI: the development of “explainable AI.”

Image segmentation is the pixel-by-pixel detection of objects, which is the most challenging but informative in the fundamental tasks of machine learning including image classification and object detection. Pixel-by-pixel segmentation is required to apply machine learning to support fetal cardiac ultrasound screening; we have to detect cardiac substructures precisely which are small and change shapes dynamically with fetal heartbeats, such as the ventricular septum. This task is difficult for general segmentation methods such as DeepLab v3+, and U-net. Hence, here we proposed a novel segmentation method named Cropping-Segmentation-Calibration (CSC) that is specific to the ventricular septum in ultrasound videos in this study. CSC employs the time-series information of videos and specific section information to calibrate the output of U-net. The actual sections of the ventricular septum were annotated in 615 frames from 421 normal fetal cardiac ultrasound videos of 211 pregnant women who were screened. The dataset was assigned a ratio of 2:1, which corresponded to a ratio of the training to test data, and three-fold cross-validation was conducted. The segmentation results of DeepLab v3+, U-net, and CSC were evaluated using the values of the mean intersection over union (mIoU), which were 0.0224, 0.1519, and 0.5543, respectively. The results reveal the superior performance of CSC.

Diagnostic support tools based on artificial intelligence (AI) have exhibited high performance in various medical fields. However, their clinical application remains challenging because of the lack of explanatory power in AI decisions (black box problem), making it difficult to build trust with medical professionals. Nevertheless, visualizing the internal representation of deep neural networks will increase explanatory power and improve the confidence of medical professionals in AI decisions. We propose a novel deep learning-based explainable representation “graph chart diagram” to support fetal cardiac ultrasound screening, which has low detection rates of congenital heart diseases due to the difficulty in mastering the technique. Screening performance improves using this representation from 0.966 to 0.975 for experts, 0.829 to 0.890 for fellows, and 0.616 to 0.748 for residents in the arithmetic mean of area under the curve of a receiver operating characteristic curve. This is the first demonstration wherein examiners used deep learning-based explainable representation to improve the performance of fetal cardiac ultrasound screening, highlighting the potential of explainable AI to augment examiner capabilities.

The application of segmentation methods to medical imaging has the potential to create novel diagnostic support models. With respect to fetal ultrasound, the thoracic wall is a key structure on the assessment of the chest region for examiners to recognize the relative orientation and size of structures inside the thorax, which are critical components in neonatal prognosis. In this study, to improve the segmentation performance of the thoracic wall in fetal ultrasound videos, we proposed a novel model-agnostic method using deep learning techniques: the Multi-Frame + Cylinder method (MFCY). The Multi-frame method (MF) uses time-series information of ultrasound videos, and the Cylinder method (CY) utilizes the shape of the thoracic wall. To evaluate the achieved improvement, we performed segmentation using five-fold cross-validation on 538 ultrasound frames in the four-chamber view (4CV) of 256 normal cases using U-net and DeepLabv3+. MFCY increased the mean values of the intersection over union (IoU) of thoracic wall segmentation from 0.448 to 0.493 for U-net and from 0.417 to 0.470 for DeepLabv3+. These results demonstrated that MFCY improved the segmentation performance of the thoracic wall in fetal ultrasound videos without altering the network structure. MFCY is expected to facilitate the development of diagnostic support models in fetal ultrasound by providing further accurate segmentation of the thoracic wall.