Scholarlink

<p>Dynamic Infrared Thermography (DIT) captures a sequence of thermal images after applying a stimulus to analyze vascular responses, offering greater sensitivity than static methods. However, its clinical application remains limited. To address this, Deep Learning (DL) models have gained attention for extracting spatial and temporal patterns automatically. In this study, we propose an explainable deep temporal modeling framework based on recurrent neural networks (RNN, LSTM, GRU) trained on dynamic thermal imaging (DTI) sequences composed of 20-frame series per patient. The architecture integrates a convolutional neural network (CNN) backbone for feature extraction, followed by Grad-CAM for interpretability, allowing the identification of regions contributing most to the classification decision. Feature selection was performed using both statistical methods (based on temporal variance) and wrapper-based optimization techniques, including genetic algorithms (GA), particle swarm optimization (PSO), and simulated annealing (SA). These subsets were used as input to the recurrent classification models. Model evaluation was performed using repeated five-fold cross-validation. All configurations achieved high classification performance, with accuracy ranging from 94% to 97%, sensitivity from 95% to 98%, and specificity from 93% to 96%. Filtering reduced dimensionality while preserving accuracy, whereas wrapper methods further boosted performance, with RNN-GA reaching 97.03% ± 0.38%. RNN models consistently outperformed LSTM and GRU variants across all subsets. The main strength of this framework lies in the integration of Grad-CAM-based feature attribution, handcrafted temporal descriptors, and deep sequence modeling, thus enhancing explainability for clinical decision-making. However, a key limitation remains the restricted dataset size.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An Explainable Deep Temporal Model of Dynamic Breast Thermography Based on Grad-CAM and a Metaheuristic Feature Selection Strategy

  • 期刊论文

  • 发表日期：

    2026年2月25日

查看全文
  • Andrés Munguía-Siu,
  • J. Horacio Espinoza-Rodríguez

摘要

Dynamic Infrared Thermography (DIT) captures a sequence of thermal images after applying a stimulus to analyze vascular responses, offering greater sensitivity than static methods. However, its clinical application remains limited. To address this, Deep Learning (DL) models have gained attention for extracting spatial and temporal patterns automatically. In this study, we propose an explainable deep temporal modeling framework based on recurrent neural networks (RNN, LSTM, GRU) trained on dynamic thermal imaging (DTI) sequences composed of 20-frame series per patient. The architecture integrates a convolutional neural network (CNN) backbone for feature extraction, followed by Grad-CAM for interpretability, allowing the identification of regions contributing most to the classification decision. Feature selection was performed using both statistical methods (based on temporal variance) and wrapper-based optimization techniques, including genetic algorithms (GA), particle swarm optimization (PSO), and simulated annealing (SA). These subsets were used as input to the recurrent classification models. Model evaluation was performed using repeated five-fold cross-validation. All configurations achieved high classification performance, with accuracy ranging from 94% to 97%, sensitivity from 95% to 98%, and specificity from 93% to 96%. Filtering reduced dimensionality while preserving accuracy, whereas wrapper methods further boosted performance, with RNN-GA reaching 97.03% ± 0.38%. RNN models consistently outperformed LSTM and GRU variants across all subsets. The main strength of this framework lies in the integration of Grad-CAM-based feature attribution, handcrafted temporal descriptors, and deep sequence modeling, thus enhancing explainability for clinical decision-making. However, a key limitation remains the restricted dataset size.