DensiThAI, A Multi-View Deep Learning Framework for Breast Density Estimation using Infrared Images

Breast tissue density is a key biomarker of breast cancer risk and a major factor affecting mammographic sensitivity. However, density assessment currently relies almost exclusively on X-ray mammography, an ionizing imaging modality. This study investigates the feasibility of estimating breast density using artificial intelligence over infrared thermal images, offering a non-ionizing imaging approach. The underlying hypothesis is that fibroglandular and adipose tissues exhibit distinct thermophysical and physiological properties, leading to subtle but spatially coherent temperature variations on the breast surface. In this paper, we propose DensiThAI, a multi-view deep learning framework for breast density classification from thermal images. The framework was evaluated on a multi-center dataset of 3,500 women using mammography-derived density labels as reference. Using five standard thermal views, DensiThAI achieved a mean AUROC of 0.73 across 10 random splits, with statistically significant separation between density classes across all splits (p « 0.05). Consistent performance across age cohorts supports the potential of thermal imaging as a non-ionizing approach for breast density assessment with implications for improved patient experience and workflow optimization.

💡 Research Summary

This paper investigates whether breast density—a key risk factor for breast cancer and a determinant of mammographic sensitivity—can be estimated from non‑ionizing infrared thermal images using deep learning. The authors hypothesize that the distinct thermophysical and physiological properties of fibroglandular (dense) and adipose (fatty) tissues produce subtle, spatially coherent temperature patterns on the breast surface that can be captured by high‑resolution long‑wave infrared cameras.

Physical rationale – Section 2 presents a steady‑state Pennes bioheat model, showing that tissue thermal conductivity, blood perfusion, metabolic heat generation, and specific heat differ markedly between glandular and adipose tissue (e.g., conductivity 0.328 vs 0.171 W/m·K). Simulations and prior experimental work suggest these differences can generate surface temperature variations on the order of 0.1–0.5 °C. Although such differences are small relative to environmental noise, they are systematic and spatially correlated, providing a theoretical basis for machine‑learning extraction.

Methodology (DensiThAI) – The proposed framework consists of four stages:

1. View Normalization – Five canonical thermal views (frontal, left/right lateral, left/right oblique) are normalized per subject using the global minimum and maximum temperature across all views, preserving relative gradients while reducing inter‑subject variability.

2. Multi‑View Feature Extraction – Each normalized view is fed into a VGG‑16‑based convolutional encoder (weights shared across views). After global average pooling, a 512‑dimensional latent vector is obtained for each view.

3. Multi‑View Fusion – A simple average pooling across the five view vectors yields a single fused representation, avoiding extra trainable parameters and mitigating over‑fitting.

4. Density Classification – A linear layer followed by a sigmoid produces a probability of “dense” breast tissue. The network is trained with binary cross‑entropy loss. Pre‑training on ImageNet and fine‑tuning on thermal data accelerate convergence.

Dataset – A multi‑center cohort of 3,500 women undergoing routine mammography provided ground‑truth density labels (binary: fatty vs. dense) derived from BI‑RADS or equivalent textual reports. Thermal images were captured with FLIR cameras (240 × 320 px, <50 mK sensitivity) under standardized environmental conditions. The cohort comprised 2,435 fatty and 1,065 dense cases, stratified by age (≤45 yr: 530 fatty / 499 dense; >45 yr: 1,905 fatty / 566 dense). Data were split into training (60 %), validation (20 %), and test (20 %) sets, and the entire pipeline was repeated over ten random seeds to assess robustness.

Results –

• Using the full‑field five‑view input, DensiThAI achieved an average AUROC of 0.73 ± 0.016 and AUPRC of 0.56 ± 0.023 across the ten splits.
• When images were cropped to the breast region only (six regions per subject), performance remained comparable (AUROC 0.72 ± 0.026, AUPRC 0.53 ± 0.040).
• Mann‑Whitney U tests on predicted scores yielded highly significant separation (p < 1 × 10⁻⁷ for full‑field, p < 1 × 10⁻¹¹ for cropped).
• Single‑view models (max‑pooling across views) performed slightly worse, confirming the benefit of multi‑view aggregation.
• A radiomics‑based baseline (first‑order + GLCM features + Random Forest) attained AUROC ≈ 0.62, markedly lower than DensiThAI, highlighting the advantage of learned deep features.
• Age‑cohort analysis showed no meaningful performance drop in either ≤45 yr or >45 yr groups, indicating that the model relies on tissue‑level thermal signatures rather than age‑related shape cues.

Conclusions and Outlook – The study demonstrates that infrared thermography, when coupled with a carefully designed multi‑view deep‑learning pipeline, can reliably discriminate dense from fatty breasts without ionizing radiation. The work contributes (1) a physics‑grounded justification for thermal density signals, (2) a lightweight yet effective architecture that leverages view‑level consistency, and (3) empirical evidence of reproducibility across centers and age groups. Future directions include tighter environmental control, incorporation of 3‑D heat‑transfer modeling, regression‑style estimation of continuous density percentages, and multimodal fusion with ultrasound or MRI to further enhance clinical utility.