Predicting Lung Cancer's Metastats' Locations Using Bioclinical Model
Teddy Lazebnik, Svetlana Bunimovich-Mendrazitsky
TL;DR
This work addresses predicting spatial locations of lung cancer metastases by integrating a biophysical framework with patient-specific 3D CT data. The proposed approach defines a mapping $M$ that combines a vascular graph, tumor location, tissue masks, and a query point to produce voxelwise metastasis probabilities, using BFS-driven vascular paths and growth decay constraints. Heatmaps are generated by evaluating the biophysical model across a voxel grid on the CT volume and normalizing with the $L_1$ metric, yielding a probabilistic forecast of metastasis locations. Validation on 10 patient cases yields around $0.74$ accuracy with low variability, illustrating the potential of physics-informed, data-efficient predictions to guide personalized diagnosis and treatment planning in lung cancer.
Abstract
Lung cancer is a leading cause of cancer-related deaths worldwide. The spread of the disease from its primary site to other parts of the lungs, known as metastasis, significantly impacts the course of treatment. Early identification of metastatic lesions is crucial for prompt and effective treatment, but conventional imaging techniques have limitations in detecting small metastases. In this study, we develop a bioclinical model for predicting the spatial spread of lung cancer's metastasis using a three-dimensional computed tomography (CT) scan. We used a three-layer biological model of cancer spread to predict locations with a high probability of metastasis colonization. We validated the bioclinical model on real-world data from 10 patients, showing promising 74% accuracy in the metastasis location prediction. Our study highlights the potential of the combination of biophysical and ML models to advance the way that lung cancer is diagnosed and treated, by providing a more comprehensive understanding of the spread of the disease and informing treatment decisions.