Related Experiment Videos
Multiclass lung cancer detection using a hybrid capsule inspired deep neural network.
Ananya Bhattacharee1, Abhishek Bhattacharjee2, Ranjit Prasad Swain3
1Symbiosis Institute of Technology, Pune Campus, Symbiosis International (Deemed University), Pune, Maharashtra, 412115, India. ananya.bhattacharjee@sitpune.edu.in.
Scientific Reports
|April 21, 2026
Summary
A new lightweight attention model improves lung cancer detection in CT scans by efficiently encoding features, achieving 100% accuracy. This approach reduces computational complexity compared to capsule networks, aiding radiologists.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Science
Background:
- Convolutional Neural Networks (CNNs) have limitations in preserving spatial relationships for lung cancer detection.
- Capsule Networks (CapsNets) address this but introduce significant computational complexity due to matrix multiplication.
Purpose of the Study:
- To propose a computationally efficient model for multiclass classification of lung CT images.
- To reduce the complexity associated with existing deep learning models for lung cancer detection.
Main Methods:
- A novel architecture combining EfficientNetB0 with a lightweight attention (LWA) module and a capsule-inspired feature encoding module.
- The LWA module filters and strengthens important features, reducing redundancy and computational load.
- A simplified matrix multiplication approach is used, resulting in low floating point operations (FLOPs).
Main Results:
- The proposed model achieved 100% F1 score, accuracy, precision, and recall on the test set.
- Demonstrated no data leakage, confirming the robustness of the results.
- Achieved a computational complexity of 0.01 GFLOPS, significantly lower than traditional CapsNets.
Conclusions:
- The proposed LWA-based EfficientNetB0 and capsule-inspired model offers a computationally efficient and highly accurate solution for lung cancer detection.
- The model shows potential to assist radiologists in diagnosing benign, normal, and malignant lung conditions from CT images.
- Further prospective clinical evaluation is needed to confirm its real-world utility.