# Deep Learning-Based Prediction of Individual Cell α-Dispersion Capacitance from Morphological Features

## TL;DR

This paper introduces a deep learning method to separate morphological effects from electrical measurements in cells, improving cancer diagnostics.

## Contribution

A novel deep learning model predicts and corrects morphology-induced capacitance errors in cell electrical measurements.

## Key findings

- A CNN model accurately predicts capacitance spectra from cell morphology with <10% error.
- The model isolates true biological responses from morphological artifacts in EGF-stimulated HeLa cells.
- The method improves EGFR-based cancer diagnostics by correcting measurement errors.

## Abstract

The biophysical characteristics of cellular membranes, particularly their electrical properties in the α-dispersion frequency domain, offer valuable insights into cellular states and are increasingly important for cancer diagnostics through epidermal growth factor receptor (EGFR) expression analysis. However, a critical limitation in these electrical measurements is the confounding effect of morphological changes that inevitably occur during prolonged observation periods. These shape alterations significantly impact measured capacitance values, potentially masking true biological responses to epidermal growth factor (EGF) stimulation that are essential for cancer detection. In this study, we attempted to address this fundamental challenge by developing a deep learning method that establishes a direct computational relationship between cellular morphology and electrical properties. We combined optical trapping technology and capacitance measurements to generate a comprehensive dataset of HeLa cells under two different experimental conditions: (i) DPBS treatment and (ii) EGF stimulation. Our convolutional neural network (CNN) architecture accurately predicts 401-point capacitance spectra (0.1–2 kHz) from binary morphological images at low frequencies (0.1–0.8 kHz, < 10% error rate). This capability allows for the identification and subtraction of morphology-dependent components from measured capacitance changes, effectively isolating true biological responses from morphological artefacts. The model demonstrates remarkable prediction performance across diverse cell morphologies in both experimental conditions, validating the robust relationship between cellular shape and electrical characteristics. Our method significantly improves the precision and reliability of EGFR-based cancer diagnostics by providing a computational framework for a morphology-induced measurement error correction.