Treating cancer is a critical challenge in modern medicine, as it affects millions of people worldwide and can have devastating effects on both patient health and quality of life. Lung cancer, in particular, is one of the most prevalent and deadly forms of cancer worldwide (1). One of the major challenges in treating lung cancer is the development of metastases, which are secondary tumors that develop from cells that have spread from the primary tumor site to other parts of the organ (or even the body) (2–4). The location of these metastases can greatly affect a patient's prognosis and the effectiveness of treatment (5).

In particular, for lung cancer diagnosis, healthcare professionals commonly use a computed tomography (CT) and Positron Emission Tomography CT (PET-CT) to diagnose the patient's clinical condition (6, 7). This non-invasive imaging modality allows for the simultaneous acquisition of functional and anatomic information, providing detailed insights into the metabolic activity of the tumor and its relationship to surrounding tissue. Additionally, this analysis provides the anatomy and structure of the lung and surrounding tissue, which can be used to determine the size and location of the tumor. Based on this data, clinicians look for the primary cancer tumor's as well as metastasis' properties to determine the course of treatment (8–10).

Recently, to partially automate the process of parsing the required clinical data from CT imagery, researchers have used computer vision and ML algorithms in general and to detect lung cancer, in particular (11–14). These algorithms can (semi-)automatically detect and classify lung nodules, reducing the dependence on human interpretation and improving the consistency of diagnoses. In addition, biophysical models gathering popularity in the clinical domain as these getting better at predicting clinical outcomes over time (15–19). For instance, Kang et al. (20) constructs a mathematical model that integrates let-7 and miR-9 expression into a signaling pathway to generate an in silico model for the process of epithelial-mesenchymal transition. The authors validate their model using in vitro data collected by testing the effects of EGFR inhibition on downstream MYC, miR-9, and let-7a expression. Hong et al. (21) propose a multicomponent mathematical model for simulating lung cancer growth as well as radiotherapy treatment for lung cancer, showing promising predictive accuracy compared to a relatively small set of in vivo data points.

Unfortunately, current PET-CT technology lacks the ability to clearly capture small metastasis, usually less than 2 mm in diameter due to the mix of cancer and healthy cells together with the noise occurring during the sampling process (22). Missing small-size metastasis can lead to choosing the sub-optimal course of treatment and result in clinically catastrophic results. In this work, we propose a bio-physical model to predict from PET-CT imagery the locations of small-size metastasis that the current PET-CT methods are missing. Data-driven models are increasingly being utilized in the field of oncology (23, 24). By utilizing patient-specific biological and clinical data, these models aim to provide a more comprehensive representation of the disease and its progression, thereby enabling more targeted and effective treatment strategies (25, 26). The use of these models in the treatment of lung cancer holds great promise, and this study aims to evaluate their ability to predict the location of lung cancer metastases, with the ultimate goal of improving patient outcomes (27).

The novelty of this work lies in the combination of three types of bio-physical models associated with cancer settlement, flow in the bloodstream, and growth inside healthy tissue. We tested the proposed model on the historical data of 10 lung cancer patients with metastasis, obtaining a 74% accuracy in the prediction of the metastasis locations.

The remainder of the paper is organized as follows. Section 2 outlines the proposed bioclinical model, divided into the biophysical model, initial condition construction from the 3D CT image, and the metastasis probability heat-map generation process. Section 3 presents the performance of the proposed model on real-world clinical data. Finally, section 4 discusses the obtained results with their clinical applications and proposes possible future work.

The spread dynamics of metastasis originating in the primary tumor can be associated with three main biological processes: (1) the flow of cancer cells in the bloodstream; (2) the settlement of cancer cells in the tissue; and 3) the spatial spread of cancer polyps (28–30). All these biological processes are spatio-temporal in nature and therefore influenced by the spatial distributions of blood vessels, healthy tissue, and the original cancer cells. To capture these settings, we use a chest PET-CT image, obtaining a three-dimensional (3D) gray-scale image (I ∈ ℝ^{x × y × z}) where x, y, z are the CT image's dimensions. Using I, and by simulating the biological and clinical occurring in vivo, one can predict the locations of cancer's metastasis, if these exist. A schematic view of the model's components and the interactions between them is summarized in Figure 1. Namely, the algorithm can be divided into three main components. First, biophysical modeling is responsible for predicting the ability of cancer cells to colonize different parts of the lungs over time. Second, the computer vision algorithm accepts the 3D CT image and produces the parameterized initial condition construction for the biophysical model. Finally, an algorithm that utilizes the biophysical model to generate the metastasis heatmap.

To evaluate the performance of the proposed bioclinical model, we used samples from 10 patients. All patients are diagnosed with metastases in the lungs in the Sheba Hospital (Israel) between 2019 and 2022. The metastasis instances have been manually tagged by a clinician, indicating the location in the lung and the size, alongside a manual tagging of the prime tumor's location and size. For these patients, a CT image before the detection of metastasis and afterward are obtained, without treatment in between, allowing for validation of the metastasis occurring locations without outside influence. Table 1 shows the socio-demographic properties of the cohort.

Table 2 outlines the result of this computation, divided into the Hard and Soft classifier scores. Overall, the bioclinical model shows 74% accuracy, on average, in predicting the location of lung metastasis from lung cancer with a 1.9% standard deviation. This indicates that the model receives stable levels of performance over the sampled population. The scores describe the accuracy metric of a Soft-classifier between two 3D images I1∈{0,1}x·y·z and I2∈ℝx·y·z where I₁ is a binary image that indicates if a cell in the image is part of a cancer polyp or not and I₂ is the prediction of the proposed model (see Section 2.3):

In addition, we formally define the Hard-classifier score to be dh:{0,1}x·y·z×{0,1}x·y·z×ℝ→ℝ such that d_h(I₁, I₂, ζ): = d_s(I₁, I₂|_ζ) where I₂|_ζ is obtained by performing a threshold ζ on the predicted image I₂. I.e. each value α ∈ I₂ is replaced with 1 if α>ζ and 0 otherwise).

One can notice that the standard deviation of both the Hard and Soft classifiers are relatively small (0.0192 and 0.0193, respectively), which indicates that the proposed model is relatively robust for this cohort. Moreover, focusing on the Hard-classifier, the performance ranges from 0.690 to 0.830 which indicates that outside parameters influence the model's performance. That said, age, gender, and smoking status are not directly correlated to the performance based on the obtained cohort.

The spread of lung cancer from its primary site to other parts of the lungs is a critical aspect of lung cancer progression and is associated with decreased survival rates (45, 46). Early identification of metastatic lesions is crucial for prompt and effective treatment, as it can significantly impact patient outcomes (47). However, conventional imaging techniques such as computed tomography (CT) scans have limitations in detecting small metastases (48, 49). In this study, we proposed a personalized biophysical-based bioclinical mathematical model that accepts a patient's 3D CT image and produces a 3D heatmap of the probability a metastasis would develop in each region of the lungs.

We tested our bioclinical model on real-world clinical data and validated it with clinician domain experts to get a baseline. As seen in Table 2, the proposed bioclinical model provides around 74% accuracy in predicting the location of metastasis. This means, that on average, three out of four predictions of the model marked locations with metastasis in the CT that would not be marked otherwise. However, these results do not provide a lower-boundary accuracy measurement of the model's performance as locations marked by the algorithm with metastasis that was without are either a wrong prediction or that metastasis was not found during the validation phase and would be created given more time. For example, let us examine a 2D slice of one of the patients in the sample. Figure 2 shows the original CT image slice (with the manual mark of the primary tumor), the blood vessels marked in blue, the healthy tissue locations, and the model's heatmap prediction - from left to right and top to bottom. One can notice that the model predicts metastasis for the right-top corner of the lungs in this z-axis value. However, metastasis is not found there. This is because either the model is mistaken or not enough time passed from the first and second CT scans in order to allow metastasis to grow into a detected size.

One of the key advantages of the bioclinical model is the fact it is based on biophysical modeling of the cancer spread which allows one to integrate diverse data sources and provides a more comprehensive understanding of the disease compared to purely data-driven methods. Theoretically, one could aim to develop a data-driven solution to this time-series task using ML or deep learning methods. However, to do so, an extensive data-gathering process of multiple CT scans of lung cancer would be required, which would be extremely expensive, logistically complex, and ethically questionable. On the other hand, the proposed approach utilizes previous biological, physical, and clinical knowledge, resulting in a need for just a relatively small dataset, used entirely to validate and evaluate the performance of the proposed model.

The results of this study demonstrate the potential of mathematical models in predicting the location of lung cancer metastases. Our findings show that the bioclinical model can provide personalized predictions of metastatic spread with decent accuracy. Hence, this model can potentially inform more targeted and effective treatment strategies for lung cancer patients, ultimately improving patient outcomes. As such, this study provides a step forward in developing more personalized and effective treatments for lung cancer. The use of bio-clinical mathematical models to predict the location of metastases holds great promise for improving patient outcomes and should be a focus of future research efforts in this field.

However, our bioclinical model can (and should) be further improved. First, the currently used biomathematical model of cancer colonization on healthy tissue does not take into consideration the spatial properties of the cancer cells. Second, cancer cells continue to mutate over time which causes a wide range of outcomes and can significantly alter the course of metastasis formation. Third, taking into consideration additional clinical data about the patient such as gender, background diseases, smoking, and others, it will be possible to obtain more accurate the location of cancer metastases (50). Forth, this study was limited by its sample size (of 10 patients) and the heterogeneity of the patient population. Thus, additional large-scale, multicentered trials are necessary to further test the accuracy and clinical utility of the proposed model, to obtain more statistically representative results. In addition, as future work, molecular dynamics simulations can be also integrated into the proposed model to further improve its accuracy (51).

Taken jointly, the proposed bioclinical model provides a new perspective of the metastasis location appearance for lung cancer using PET-CT data. This model can be medically useful for patients with stage two lung cancer where it is not clear from the PET-CT image itself if metastasis already occurred and is too small to detect or has not occurred yet by highlighting potential locations in the image for medical experts, allowing a later medical decision process with better information. Nonetheless, further research should be made to validate the obtained results on a larger cohort to statistically establish its usefulness.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

TL: Conceptualization, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Visualization, Writing – original draft, Writing – review & editing. SB-M: Data curation, Writing – original draft, Writing – review & editing.