Lightweight and explainable CNNLSTM model for accurate lung cancer detection in thoracic CT images

To have a positive impact on patient survival and death rates, lung cancer must be diagnosed early and as accurate as possible. While deep learning has shown exceptional success in medical image analysis, a large number of the current methods use computationally intensive architectures that are not very interpretable and thus cannot be used in resource-limited healthcare settings. In this paper, a lightweight and explainable deep learning framework for automatic lung cancer diagnosis from computed tomography (CT) images of the thorax is proposed. The proposed framework uses Gaussian filtering and Contrast Limited Adaptive Histogram Equalization (CLAHE) to improve the quality of the image before it is segmented using a pre-trained U-Net model for lung segmentation to accurately isolate lung regions. Discriminative spatial features are extracted using MobileNetV2 network, and then processed by a Long Short-Term Memory (LSTM) network to learn the relations between spatial feature representations. The class imbalance issue is mitigated by the addition of Binary Focal Loss during optimization of the model, and the visual model explanations using Gradient-weighted Class Activation Mapping (Grad-CAM). The framework is tested on the publicly available LIDC-IDRI dataset using the train validation test stratification approach based on patient. Experimental results show that the proposed approach can reach an accuracy of 94.8%, and the precision, recall, F1 score and ROC–AUC are competitive classification accuracy metrics, while being computationally efficient for edge deployment. Additionally, Grad-CAM visualizations add to the transparency of the model by identifying clinically relevant areas that aid in the diagnosis. While multicenter validation is still to be done outside the lab, the proposed framework has a favourable trade-off of diagnostic accuracy, low computational requirements, and high explainability, promising it as a decision-support system for computer-aided lung cancer diagnosis.