Cauchy Lotus Optimization-based feature selection and ResNet 101 based XceptionNet architecture for radiation pneumonitis prediction in lung cancer patients.
Authors
Affiliations (6)
Affiliations (6)
- Department of Future Technology, Korea University of Technology and Education, Cheonan, South Korea. Electronic address: [email protected].
- Department of Biomedical Engineering, Vel Tech Rangarajan Dr.Sagunthala R&D Institute of Science and Technology, Avadi, Tamil Nadu, India. Electronic address: [email protected].
- Department of Electronics and Computer Engineering, CSMSS Chh. Shahu College of Engineering, Chhatrapati Sambhajinagar, Maharashtra, India. Electronic address: [email protected].
- Department of Computer Science and Engineering, Faculty of Engineering, Karpagam Academy of Higher Education (Deemed to be University), Coimbatore, Tamil Nadu, India.
- Department of UdalThathuvam - Human Physiology, Sri Sairam Siddha Medical College and Research Centre, Chennai, Tamil Nadu, India.
- Annai Vailankanni College of Engineering, Kanyakumari, Tamil Nadu, India. Electronic address: [email protected].
Abstract
Thoracic radiotherapy for lung cancer patients followed by radiation pneumonitis (RP) has significant clinical side effects. Risk-adaptive treatment planning can be supported by accurate early RP prediction. Using thoracic CT scans, this study suggests an efficient deep learning algorithm for RP prediction. An analysis was conducted on a retrospective cohort of 548 patients with lung cancer who received thoracic radiotherapy between 2010 and 2021. According to established toxicity criteria, clinically significant RP was classified as Grade ≥ 2 and evaluated during post-treatment follow-up. Clinically accessible radiation outlines were used to separate bilateral lung regions, and an improved ResNet101-based XceptionNet architecture was used to extract deep features from CT images. Cauchy Lotus Optimization (CLO) was used for feature selection in order to minimize redundancy after an autoencoder was used for compact feature representation. At the patient level, the dataset was divided into cohorts for independent training (80%) and testing (20%). To avoid information leaking, only the training data was used for feature selection and model training. Precision, specificity, sensitivity, accuracy, and ROC-AUC were used to assess performance on the independent test set. Emperor Penguins Colony Algorithm (EPCA), Sea Lion Optimization (SLO), Spotted Hyena Optimization (SHO), Marine Predator Optimization (MPO), and other optimization-based techniques were compared to representative deep learning baselines. On the independent test set, the suggested framework demonstrated excellent predictive performance with high precision, accuracy, sensitivity, and specificity. Strong discriminative ability was shown by ROC analysis, and the suggested approach produced the highest AUC when compared with competing techniques. While maintaining discriminative power, the optimized feature selection technique significantly decreased feature dimensionality.