Automated Urine Cytopathology Reporting System

Greatly reduces the heavy workload on pathologists by assisting with clinical decision-making: developing deep learning models for cell-level analysis, integrating clinical knowledge and statistical shape priors, and using cell-level knowledge and whole-slide information to infer diagnosis results.

• Novel knowledge-regulated Convolutional Neural Network (CNN) designed for the segmentation of cellular regions from whole slide images. Integrates simple yet important prior knowledge into a segmentation network, which helps avoid unreasonable segmentation errors.
• Deep learning model capable of learning and updating the shape priors of cytoplasm from feature maps generated during its training procedure. Designed to segment overlapping cytoplasm by effectively leveraging shape priors to guide the training and refine the segmentation results.
• Consists of a novel CNN that includes feature embedding and sample imbalance learning schemes designed to improve cell classification accuracy. Also includes a diagnosis results prediction model that uses a multi-class U-statistic approach.

• Will help clinical experts reduce examination times, boost diagnosis accuracy of bladder cancer, and reduce mortality rate of bladder cancer.
• Will benefit and support companies working on medical image analysis, medical device manufacturers, and medical service providers of intelligent clinical systems.
• Has the potential to enhance research on deep learning and its relevant industries by providing valuable insights on how to construct deep learning models with good capacity and generality. It achieves this by leveraging both data and prior knowledge.
• Has great potential to benefit the healthcare industry, particularly companies focusing on providing automated intelligent diagnosis and clinical decision-making.

• Automated Bladder Cancer Diagnosis