Adaptive Density-Informed AI Framework for Improving Breast Cancer Detection Using Mammography and Thermal Imaging - 14/04/26
, Bharath Govindaraju 1, Sudhakar Sampangi 2, Geetha Manjunath 1
Bienvenue sur EM-consulte, la référence des professionnels de santé.
Article gratuit.
Connectez-vous pour en bénéficier!
Highlights |
• | Introduces a Density-Informed Multimodal AI (DIMA) framework for breast cancer screening. |
• | Proposes a clinically feasible AI system combining mammography and thermal imaging to overcome dense-breast limitations. |
• | Achieves 94.6% sensitivity and 79.9% specificity in a cohort of 324 women across both dense and fatty breast types. |
• | Employs an interpretable, rule-based fusion using breast density as a conditional decision variable. |
• | Enables a clinically deployable, low-cost, and equitable AI pathway for breast cancer detection. |
Abstract |
Background |
Mammographic screening performance declines in women with dense breasts, leading to diagnostic inequities and delayed cancer detection. Thermal imaging captures complementary functional cues such as vascular and metabolic activity that remain detectable irrespective of density with the help of Artificial Intelligence (AI). This study presents an adaptive Density-Informed Multimodal AI (DIMA) Framework that integrates mammography and thermal imaging to improve breast cancer detection performance across breast density categories.
Methods |
The framework integrates two complementary AI pipelines: a multi-view deep learning model trained on 19,883 mammograms for morphological feature analysis, and a radiomics-based Thermalytix system trained on over 100,000 thermal images to capture vascular and thermal asymmetries. Breast density, categorized by ACR grades, functions as a conditional variable that dynamically determines which model’s prediction is used. The mammography AI is applied for fatty breasts (ACR A and B), whereas the Thermalytix AI is prioritized for dense breasts (ACR C and D). This density-conditioned decision logic enables optimal modality utilization while maintaining interpretability. To assess real-world applicability, the framework was evaluated on 324 women who underwent both mammography and thermal imaging.
Results |
The DIMA framework achieved a sensitivity of 94.6% (95% CI: 88.6–100) and specificity of 79.9% (95% CI: 75.1–84.7), outperforming standalone mammography AI (sensitivity 81.8%, specificity 86.3%) and Thermalytix AI (sensitivity 92.7%, specificity 75.5%). Importantly, the sensitivity of Mammography dropped significantly in dense breasts (67.9%) versus fatty breasts (96.3%), whereas Thermalytix AI maintained high and consistent sensitivity in both (92.6% and 92.9%, respectively).
Conclusions |
This retrospective cross-sectional diagnostic accuracy evaluation demonstrates the potential of a DIMA routing strategy to improve breast cancer detection across breast density categories. Population-based screening studies are required to assess its generalizability, equity, and role within large-scale screening programs.
Le texte complet de cet article est disponible en PDF.Keywords : Breast Cancer, Mammography, Thermalytix, Thermal Imaging, Multi-modal imaging, Artificial Intelligence, Machine Learning
Plan
| This study has not been presented in any other journal. |
|
| Conflict of interest |
|
| Siva Teja Kakileti reports a relationship with NIRAMAI Health Analytix Pvt. Ltd. that includes employment and equity ownership. Bharath Govindaraju reports a relationship with NIRAMAI Health Analytix Pvt. Ltd. that includes employment. Sudhakar Sampangi reports a relationship with NIRAMAI Health Analytix Pvt. Ltd. that includes consulting or advisory roles and equity ownership. Geetha Manjunath reports a relationship with NIRAMAI Health Analytix Pvt. Ltd. that includes board membership, employment, and equity ownership. |
|
| Funding Support |
|
| This study was funded internally by Niramai Health Analytix Pvt Ltd. No external funding was received. The funding source had no role in standard-of-care interpretation, clinical decision-making, or patient management. AI analyses were performed offline and independently of clinical reporting. |
|
| Author Contributions |
|
| Siva Teja Kakileti: Conceptualization, Methodology, Investigation, Formal Analysis, Software, Visualization, Writing – Original Draft, Supervision. |
|
| Bharath Govindaraju: Methodology, Formal Analysis, Data Curation, Software, Visualization, Writing – Original Draft. |
|
| Sudhakar Sampangi: Conceptualization, Resources, Writing – Review & Editing. |
|
| Geetha Manjunath: Conceptualization, Resources, Supervision, Writing – Review & Editing. |
Bienvenue sur EM-consulte, la référence des professionnels de santé.