A term weighting approach exploiting external data for cancer clause classification from free-text radiology reports

Abstract

Radiology reports are written by a medical expert via analysing radiology images such as CT and MRI. It consists of cancer clause and non-cancer clauses. We focus on text classification for cancer and non-cancer classes. This data has two unique characters. First, the number of cancer clauses is much smaller than the number of non-cancer clauses. Second, important terms for cancer also occur in the non-cancer class. Since it is often difficult to determine the cancer based on radiology images, some clauses are labelled as non-cancer in spite of having important terms for cancer. Recently, term weighting approaches have been proposed to solve the data imbalance problem. However, we argue that it sometimes gives weight wrongly due to duplicate terms. Consequently, we utilize cancer related external data to calculate term weights. Since external data is highly related with cancer, we can find important terms for cancer and calculate its weight. Based on calculated weights from external data, term weights in the cancer class are increased and term weights in the non-cancer class are decreased. Through the experiment, proposed method showed enhanced performance than term weighting methods using the training data.