Automated lung cancer classification using intensity-driven RoI selection and transfer learning.
Authors
Affiliations (5)
Affiliations (5)
- School of Computer Science and Engineering, REVA University, Bengaluru, Karnataka, India. [email protected].
- School of Computer Science and Engineering, REVA University, Bengaluru, Karnataka, India.
- Department of CSE (AI&ML), CVR College of Engineering, Hyderabad, India.
- School of Computer Science Engineering and Technology, Bennett University, Greater Noida, Uttar Pradesh, 201310, India.
- School of Computer Science Engineering and Technology, Bennett University, Greater, Noida, India.
Abstract
Lung cancer diagnosis increasingly relies on advanced medical imaging systems and expert interpretation across heterogeneous clinical data sources. However, accurate clinical decision-making remains challenging due to variations in expert judgment and the complexity of extracting discriminative patterns from electronic health records and radiological datasets. To address these challenges, this paper proposes an enhanced lung cancer classification framework that leverages intensity-driven region-of-interest (RoI) selection from publicly available benchmark datasets, including the Lung Image Database Consortium Image Collection (LIDC-IDRI) and The Cancer Imaging Archive (TCIA). The proposed methodology incorporates customized label refinement and precise annotation of vulnerable RoI regions to capture clinically relevant features associated with malignant nodules. A high-dimensional RoI mapping strategy is employed to improve feature representation and discrimination. Furthermore, a feedback-driven optimization mechanism is integrated within a transfer learning framework to iteratively refine model parameters and enhance learning stability. The optimized RoI representations are transferred to customized deep learning models, enabling efficient knowledge reuse and robust decision-making. The proposed approach is implemented using the CoVNet architecture and evaluated under a 60:40 training-testing split. Experimental results demonstrate a classification accuracy of 97.84%, validating the effectiveness of the proposed framework in improving predictive performance for lung cancer classification.