Integrate Sequence Information of Dose Volume Histogram in Training LSTM-based Deep Learning Model for Lymphopenia Diagnosis

Abstract

We previously demonstrated that the mean doses of lung, heart, thymus, and thoracic duct are significant for radiation induced lymphopenia. Here we hypothesized that deep learning model based on LSTM can integrate the full dose volume histogram (DVH) data and the clinical information to predict lymphopenia more accurately in patients with non-small cell lung cancer.Patients with primary lung cancer treated with definitive or palliative radiotherapy, with pre and post with full DVH data and clinical information were eligible. Lymphopenia was defined as absolute counts. Patients were randomly split as train dataset containing 83 patients and validation dataset containing 21 patients. The models tested consisted Bi-directional LSTM neural networks, LSTM neural networks, and full connected networks, which can output the lowest level of lymphocyte count during radiotherapy and the level of the largest change in lymphocyte count compared with initial lymphocyte count at the same time. Additionally, the analysis was compared with random forest method with simple features of stage, chemotherapy, mean dose of lung, heart, thymus, and thoracic duct.A total of 104 patients were eligible (age median = 62.00 years). 83 out of 104 had lymphopenia at the end of radiation therapy. There were 84 patients that were male and 20 that were female. The number of patients in staged I, stage II, stage III and stage IV were 5, 5, 64, and 30, respectively. LSTM-based Deep Learning Model was generated in the training dataset with area under the receiver operating characteristic curve (AUC) of 0.80, and validated on validation dataset with AUC of 0.80. The AUC from the random forest model of the clinical factors plus simple mean doses of the organs were 0.62.This study suggests that the LSTM-based deep learning model of integrating the full dosimetric information with the clinical information to improve lymphopenia predictions in patients with lung cancer treated with radiotherapy. Independent validation is needed before this model can be used to guide clinical practice to reduce radiation lymphopenia.