Navigating Malignancy: Deep Learning for Mediastinal Lymph Node Classification

Due to the complex anatomical placements of the chest’s mediastinal lymph nodes (LNs), differentiating between malignant and benign LNs requires invasive pathological testing, which is a complicated and unpleasant process. Malignancy identification has already been explored using image-based automatic machine-learning approaches. However, these conventional methods are hampered by the intricate selection of hand-crafted features and trade-offs between performance factors. Deep learning systems surpass standard machine learning methods and can circumvent these difficulties. However, the performance of the existing Deep Convolutional Neural Network (DCNN)-based models is inadequate due to the unavailability of high-resolution Computer Tomography (CT) pictures and an appropriate classification strategy. This research proposes a customized DCNN to solve the problem of mediocre performance. Advanced feature extraction approaches, Leaky ReLU activation to resolve dying ReLU difficulties, and RMSProp optimization have been added to the proposed DCNNs. From the TCIA public dataset, 0.36 million cancer and non-cancer CT images were acquired to pre-train the ImageNet weight and achieve the best weight. Then, the network was trained using 271 mediastinal LNs, comprising 133 malignant and 138 benign LNs. Five DCNN models (VGG19, VGG16, ResNet50, InceptionV3, and Xception) were constructed with convolution layers, max pool layers, flatten layer, fully connected layers, feature extraction layer, and SoftMax activation. The performance was compared to cutting-edge techniques. The modified VGG16 model with bottleneck feature extraction classified benign and malignant LNs with 98.08% accuracy.