Videl

Abstract

Leukemia is one of TOP 10 cancers in China, especially to young children and elderly. Treating Leukemia at the early stage can significantly increase the cure rate. However, detecting early stage Leukemia, i.e, finding a small amount of blood cells among massive normal cells, is both technically challenge and labor intensive. Therefore, we designed an AI-powered diagnoser for precise identification of white blood cells. The precise identification has two steps, blood image collecting and cell classification. First, with the help of doctors in Ruijing Hospital, we have been collecting about 1,000 high-quality labeled blood cell samples each week and accumulated more than 200,000 samples in total. This is the largest blood cell image dataset for leukemia in China. Second, based on these blood cell samples, we adopted the ResNet-variant classification method and achieved above 95% accuracy in 17 subtypes of white blood cells. The classification accuracy is comparable to experienced doctors in Ruijing Hospital. Third, models are trained on the latest Intel Knights Landing 7210 Platform with Intel Omni-Path interconnections. Applying MKL-optimized matrix operatiors and RDMA-enable comunication library (PSM) shortens the training time from more than 6 hours to less than an hour. Our AI-powered diagnoser can reduce the time and the cost of early Leukemia diagnosis from one week to one day, and 1,000 USD to 100 USD, respectively.