Explainable Lymph Node Diagnosis Model Combining MobileNet and ConvNextTiny

Computer-Aided Diagnostic (CADx) systems have proven effective in classifying pulmonary nodules. However, these models’ reliability issue represents a topic of research and discussion. This paper proposes an interpretable lightweight hybrid model for diagnosing lymph nodes using computed tomography images and deep learning. The hybrid model combines MobileNetV3Small with ConvNextTiny to highlight features and improve performance by combining two datasets to assess data diversity, including LIDC-IDRI and the chest CT scan images for the lung cancer dataset. The model is evaluated using an external dataset to show its generalization capability. LIME is used to explain the model’s decisions. The approach achieved a low false positive rate (FPR) and false negative rate (FNR) of 1.04%, resulting in high performance in all the metrics with an accuracy of 98.56% and a receiver operating characteristic (ROC) of 100% in the IQ-OTH/NCCD dataset.