One-Staged Attention-Based Neoplasms Recognition Method for Single-Channel Monochrome Computer Tomography Snapshots

Abstract

Computer tomography is most commonly used for diagnosing lung cancer, which is one of the deadliest cancers in the world. Online services that allow users to share their single-channel monochrome images, in particular computer tomography scans, in order to receive independent medical advice are becoming wide-spread these days. In this paper, we propose an optimization for the previously known two-staged architecture for detecting cancerous tumors in computer tomography scans that demonstrates the state-of-the-art results on Open Joint Monochrome Lungs Computer Tomography dataset. Modernized architecture allows to reduce the number of weights of the neural network based model (4 920 073 parameters vs. 26 468 315 in the original model) and its inference time (0.38 vs. 2.15 s in the original model) without loss of neoplasms recognition quality (0.996 F1 score). The proposed results were obtained using heavyweight encoder elimination, special combined loss function and watershed based method for the automated dataset markup and a consistency regularization approach adaptation that are described in the current paper.