Lung cancer growth forecasting and classification models by machine learning using a novel 2D co-culture high throughput device

Abstract

**Please note that the full text is embargoed until 02/01/2025** Co-culture devices have commonly been used during in vitro experimentation to create therapeutic testing devices for lung cancer that more closely mimic the in vivo environment to gain better insights into new therapies and the reasons for inconsistency for current therapies. Moreover, AI (artificial intelligence) in the form of convolutional neural networks (CNN) and convolutional long-short term memory neural networks (convLSTM) are being used to classify, diagnose, and predict drug outcomes from the visual data provided by clinical imaging modalities, such as CT-scans, MRIs, or histopathological slides among others, or from numerical and textual data gathered from the genetic testing that has already improved clinical outcomes in the form of personalized medicine. Lung cancer is often found at late stages, so it is crucial to classify the subtype and then select the most effective therapies available for the patient in a timely manner, especially if the solutions can supplement the current standards of treatment without taxing the limited resources such as biopsy tissue. With these challenges in mind, I have created a 2D co-culture device that combines lung cancer and fibroblasts to study drug effects on the combination of cells in a high throughput manner, that is easily imaged due to the island of cancer cells formed in the center of the fibroblasts. When studying the high throughput drug screening I noticed that the cancer may be growing in patterns, so I developed a CNN model that, at this early stage, can classify between two cell lines (and subtypes) of lung cancer, A549 and H460 as early as 2 and 3 days of outgrowth. Beyond this, I co-developed a convLSTM model that can forecast lung cancer outgrowth over fibroblasts utilizing the same 2D co-culture device. The machine learning model showed accurate reproduction of images matching the ground truth images and currently generated 5 to 10 days of forecasting images that agree with the growth extrapolation of the ground truth images.