Related Experiment Videos
Pulmonary CT classification based on CNN and Grad-CAM explainability using VGG16 and MobileNetV2
Chenteng Chen1, Zilong Li1, Runnan Chen1
1Department of Respiratory and Critical Care Medicine, The First Hospital of Putian City, Fujian Province, Putian, China.
Objectives:
This study compared the classification performance of VGG16 and MobileNetV2 models for normal-versus-lung-cancer-positive pulmonary CT image classification, with Grad-CAM used for post-hoc visual explainability.
Methods:
A hospital-provided dataset of 550 non-contrast thoracic CT images, comprising 240 normal and 310 lung cancer-positive images, was divided into training, validation and test subsets using a patient-grouped stratified 50/10/40 split. Images were acquired at 120 kV with a 512 × 512 matrix using a NeuViz 128 CT scanner. Both models used the same preprocessing, training-only augmentation and evaluation protocol. The VGG16 model used stacked convolutional blocks, whereas MobileNetV2 used inverted residual bottlenecks with global average pooling, dropout and a sigmoid classification layer. Performance was assessed using accuracy, sensitivity, specificity, precision, F1-score, balanced accuracy, ROC-AUC, MCC and Cohen's kappa, while Grad-CAM examined decision-relevant regions.
Results:
On the 220-image independent test set, VGG16 achieved 0.9636 accuracy, 0.9677 sensitivity, 0.9583 specificity and 0.9569 ROC-AUC. MobileNetV2 achieved 0.9909 accuracy, 0.9919 sensitivity, 0.9896 specificity and 0.9858 ROC-AUC, with fewer false-positive and false-negative predictions.
Conclusions:
MobileNetV2 achieved stronger held-out test performance than VGG16 under identical experimental conditions. Larger multi-centre patient-level validation remains necessary before clinical deployment in CT screening workflows.