Classification of body mass index levels using breast thermography: A preliminary proof-of-concept analysis with convolutional neural networks

Abstract

Background

Thermal imaging has shown promise in distinguishing between obese and non-obese individuals, yet its potential in stratifying thinner body mass index (BMI) categories remains largely unexplored.

Methods

We utilized thermal images depicting the upper abdomen to the neck of women (aged 18–81) with benign breast pathology, each comprising anterior, oblique left, and oblique right views. Employing transfer learning and convolutional neural networks (CNN), we classified images into normal weight, overweight, and obesity categories. GradCAM activation maps identified influential areas in the images.

Results

84 women were included in the analysis, with a mean age of 45 years (standard deviation (SD): 12.2). The average BMI was 28.6 kg/m² (SD: 6.5), and BMI categories were evenly distributed. The overall accuracy was 87.0 %. The model performed best for the obesity category (precision: 100 %, recall: 93 %, and F1 score: 97 %). The lowest precision was observed for the normal weight category (75 %), while the overweight category had the lowest recall (67 %) and F1 score (76 %). The confusion matrix showed that misclassifications were predominantly between the normal weight and overweight categories. Activation maps showed that in the normal weight group, the sternum area was most influential; for the obesity group, regions above the armpits were emphasized; and for the overweight group, key areas included the upper abdomen, below the breasts, and upper chest.

Conclusions

Thermal imaging effectively differentiated between BMI categories. Further validation of thermal imaging's predictive capabilities is warranted considering the limited and women-only sample herein analyzed.