AI-Driven Nonlinear Optimization Model for Early Lung Tumor Growth Prediction Using CT Imaging and Machine Learning Algorithms

Abstract

The timely detection of lung tumour development is important in the planning of individual therapy and enhancing survival. Although deep learning models have demonstrated good performance in tumour analysis based on CT, their low interpretability inhibits their integration in clinical practise. This paper presents a hybrid using AI and nonlinear tumour growth modelling and machine learning optimisation to forecast the initial progression of lung tumours based on CT scan data. The growth model of a nonlinear set of nonlinear differential equations (Gompertz /logistic) is fitted by first constrained nonlinear optimization on the sequential tumour volumes. The estimated parameters are the growth rate (r), carrying capacity (K), as well as the deceleration factor, then they are incorporated into a deep convolutional network, which will be trained to predict tumour size at subsequent time points. A longitudinal CT dataset of lung cancer patients at an early stage (N = 52) was experimented on to show that the hybrid model reduced prediction RMSE to 4.23 cm³, which was better than base line CNN-only and purely mechanistic models by 45.9% and 34.0% respectively. There was also a significant correlation between growth parameters and clinical progression (p < 0.01) thus increasing the interpretability of the models. The suggested framework is efficient in closing the gap between mechanistic nonlinear modelling and current deep learning and leads to robust, interpretable, and clinically meaningful predictions of tumour growth