Abstract: A multi-functional, computer-aided gastroscopy system optimized with integrated AI solutions is disclosed. The system makes use of multiple deep-learning neural models to achieve low latency and high-performance requirements for multiple tasks. The optimization is made at three levels: architectural, modular and functional level. At architectural level, the models are designed in such a way that it is able to accomplish HP infection classification and detection of some lesions for one inference in order to reduce computation costs. At modular level, as a sub-model of HP infection classification, the site recognition model is optimized with temporal information. It not only improves the performance of HP infection classification, but also plays important roles for lesion detection and procedure status determination. At functional level, the inference latency is minimized by configuration and resource aware optimization.

Abstract: A multi-functional, computer-aided gastroscopy system optimized with integrated AI solutions is disclosed. The system makes use of multiple deep-learning neural models to achieve low latency and high-performance requirements for multiple tasks. The optimization is made at three levels: architectural, modular and functional level. At architectural level, the models are designed in such a way that it is able to accomplish HP infection classification and detection of some lesions for one inference in order to reduce computation costs. At modular level, as a sub-model of HP infection classification, the site recognition model is optimized with temporal information. It not only improves the performance of HP infection classification, but also plays important roles for lesion detection and procedure status determination. At functional level, the inference latency is minimized by configuration and resource aware optimization.

Abstract: An image volume formed by plural anatomical images each having plural image slices of different imaging modalities is segmented by a 2D convolutional neural network (CNN). An individual anatomical image is preprocessed to form a mixed-context image by incorporating selected image slices from two adjacent anatomical images without any estimated image slice. The 2D CNN utilizes side information on multi-modal context and 3D spatial context to enhance segmentation accuracy while avoiding segmentation performance degradation due to artifacts in the estimated image slice. The 2D CNN is realized by a BASKET-NET model having plural levels from a highest level to a lowest level. The number of channels in most multi-channel feature maps of a level decreases monotonically from the highest level to the lowest level, allowing the highest level to be rich in low-level feature details for assisting finer segmentation of the individual anatomical image.

Abstract: An image volume formed by plural anatomical images each having plural image slices of different imaging modalities is segmented by a 2D convolutional neural network (CNN). An individual anatomical image is preprocessed to form a mixed-context image by incorporating selected image slices from two adjacent anatomical images without any estimated image slice. The 2D CNN utilizes side information on multi-modal context and 3D spatial context to enhance segmentation accuracy while avoiding segmentation performance degradation due to artifacts in the estimated image slice. The 2D CNN is realized by a BASKET-NET model having plural levels from a highest level to a lowest level. The number of channels in most multi-channel feature maps of a level decreases monotonically from the highest level to the lowest level, allowing the highest level to be rich in low-level feature details for assisting finer segmentation of the individual anatomical image.

Abstract: In cancer-cell screening, a patient's cells are classified by a convolutional neural network (CNN) to identify abnormal cells. In one approach, a mask having a center more transparent than the mask's periphery is used to mask an input image containing a cell of interest to yield a masked image. Since the cell is usually located around an image center, and since the image often contains irrelevant objects, such as normal cells and micro-organisms, around an image periphery, interference due to the irrelevant objects in training the CNN and in classification is diminished by using the masked image rather than the original one. In another approach, masking is applied to feature maps before classification. In the CNN, this masking is accomplished by convolving each feature map with a convolutional kernel to produce an intermediate feature map followed by chopping off a peripheral region thereof to yield a downsized feature map.