Dynamic Thermography-Based Early Breast Cancer Detection Using Multivariate Time Series.

Area of Science:

• Computational imaging
• Biomedical engineering
• Oncology

Background:

• Early breast cancer detection remains a critical challenge in women's health.
• Non-invasive screening methods are needed to improve patient outcomes and reduce healthcare costs.
• Dynamic Infrared Thermography (DIT) offers a potential non-contact method for assessing breast tissue physiology.

Purpose of the Study:

• To develop and evaluate a computational approach for early breast cancer detection using Dynamic Infrared Thermography (DIT).
• To investigate the correlation between dynamic thermal responses and tumor characteristics like angiogenesis and metabolic activity.
• To assess the efficacy of machine learning models in classifying breast cancer from DIT data.

Main Methods:

• Thermograms were acquired over 20 time points under cold-stress conditions.
• Multivariate time series features (temperature, heterogeneity, heat flux, depth) were extracted from thermal hotspots.
• Feature estimation utilized the inverse solution of the Pennes bio-heat equation, followed by classification with Time Series Forest (TSF) and Long Short-Term Memory (LSTM) networks.

Main Results:

• The Time Series Forest (TSF) model achieved an accuracy of 86% in breast cancer classification.
• The Long Short-Term Memory (LSTM) network demonstrated superior performance, reaching 94% accuracy.
• Dynamic thermal patterns effectively reflected underlying tumor angiogenesis and metabolic activity.

Conclusions:

• The combination of multivariate thermographic sequences, biophysical modeling, and machine learning shows significant potential for non-invasive breast cancer screening.
• DIT, analyzed computationally, can provide valuable insights into tumor physiology for early detection.
• The high accuracy achieved by the LSTM model highlights its suitability for analyzing complex dynamic thermal data in oncology.