End-to-end Multiple Instance Learning for Whole-Slide Cytopathology of Urothelial Carcinoma
As a non-invasive approach, cytopathology of urine sediment is a highly promising approach to diagnosing urothelial carcinoma. However, computational assessment of the cytopathological status of a sample raises the challenge of identifying few cancerous cells among thousands of cells in a microscopic whole-slide image. To address this challenge, we propose an end-to-end trainable multiple instance learning approach that combines the attention mechanism and hard negative mining to classify hematoxylin and eosin stained patient-level whole-slide images of urine sediment cells. The singular cells are extracted by a simple foreground detection algorithm. With feature embeddings computed for each image patch in a bag by a convolutional neural network, the attention mechanism serves as the pooling operator, enabling a bag-level prediction while still giving an interpretable score for each image patch. This enables the identification of key instances and potential regions of interest that trigger a patient-level decision. Our results show that the proposed system can differentiate between normal and cancerous urothelial cells, thus enabling the non-invasive diagnosis of urothelial carcinoma in patients using urine sediment analysis.
PDF Abstract
