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 fully convolutional network (FCN) implemented on a specialized processor optimized for convolution computation can achieve a speed-up in cell classification. Without re-optimizing the specialized processor, a further speed-up is achieved by compacting a testing image of cells, and processing the compacted testing image with the FCN. The testing image is first segmented into a background and regions of interest (ROIs). The ROIs are packed closer together by rearranging the ROIs without resizing them under a constraint that any two adjacent rearranged ROIs are separated by a distance in pixel not less than a minimum distance determined according to stride values of FCN convolutional layers. Geometrical operations in ROI rearrangement include relocating the ROIs and, optionally, rotating the ROIs. The rearranged ROIs are enclosed by a boundary, typically a rectangular boundary, to form the compacted testing image having an area smaller than that of the testing image.

Abstract: A fully convolutional network (FCN) implemented on a specialized processor optimized for convolution computation can achieve a speed-up in cell classification. Without re-optimizing the specialized processor, a further speed-up is achieved by compacting a testing image of cells, and processing the compacted testing image with the FCN. The testing image is first segmented into a background and regions of interest (ROIs). The ROIs are packed closer together by rearranging the ROIs without resizing them under a constraint that any two adjacent rearranged ROIs are separated by a distance in pixel not less than a minimum distance determined according to stride values of FCN convolutional layers. Geometrical operations in ROI rearrangement include relocating the ROIs and, optionally, rotating the ROIs. The rearranged ROIs are enclosed by a boundary, typically a rectangular boundary, to form the compacted testing image having an area smaller than that of the testing image.