RL2NdgsNet: Reinforcement learning based efficient classifier for mediastinal lymph nodes malignancy detection in CT images

Diagnosis of lymph nodes (LNs) in benign and malignant categories significantly impact the treatment planning of cancer. Invasive methods like biopsy are complex and painful procedures for cancer detection. Radiological modalities (like X-ray, USG, CT, and MRI) based medical imaging approaches are non-invasive and painless procedures for the cancer detection. Deep learning (DL) based architectures significantly improved the performance of medical imaging related tasks like organ detection and disease diagnosis. This paper introduces the reinforcement learning (RL) based DL network named – RL2NdgsNet (reinforcement learning based lymph nodes diagnosis network), for diagnosis of mediastinal lymph nodes (MLNs) in benign and malignant categories. Furthermore, we have designed custom DQN policy for the proposed RL2NdgsNet and also experimented with various state-of-the-art activation functions and exploration fraction. The proposed RL2NdgsNet achieves best results of sensitivity = 98.03%, specificity = 98.35%, accuracy = 98.2%, and AUC = 98.19%.