Portable Breast Mass Detection System Driven by Multimodal Deep Learning

This article presents an innovative portable breast health screening device that integrates two deep learning branches—thermal imaging CNN and pressure sensor autoencoder—to enable early automatic detection of breast masses. It aims to provide a low-cost, easily deployable breast cancer screening solution for resource-scarce areas. The system combines complementary information from thermal imaging (capturing temperature abnormalities) and pressure sensing (identifying tissue hardness), achieving detection accuracy close to professional equipment on edge AI platforms.

Breast cancer is the most common malignant tumor among women globally, and early detection is crucial for improving cure rates. However, in developing countries and remote areas, professional breast examination equipment (such as mammography) is expensive, bulky, and requires professional operation and interpretation, leading to many women being unable to undergo timely screening. Technological innovation provides a new path to address this health inequality issue.

Thermal Imaging Mechanism: High metabolic activity of tumors leads to local temperature elevation (malignant areas are 1-2°C higher than normal). Infrared sensors can non-invasively identify abnormalities, with advantages of no radiation and painlessness, but are prone to false positives due to environmental/physiological factors.

Pressure Sensing Mechanism: Malignant tumors are harder; pressure sensor arrays record deformation and reaction forces during pressing, analyzing pressure distribution to identify hardness abnormalities. The two modalities complement each other, laying the foundation for multimodal fusion.

Hardware Platform: High-resolution infrared thermal imaging module, flexible pressure sensor array, embedded computing platform (Raspberry Pi/Jetson Nano), integrated into a handheld casing, battery-powered, weighing a few hundred grams, suitable for grassroots scenarios.

Dual-Branch Neural Network: The thermal imaging branch uses CNN (to extract spatial patterns), and the pressure branch uses autoencoder (to process high-dimensional spatiotemporal sequences).

Feature Fusion: The outputs of the two branches are concatenated into a joint feature, which is input to the classification layer to obtain healthy/abnormal results, balancing accuracy and efficiency.

Dataset Construction: Collaborate with medical institutions to obtain desensitized data, generate synthetic data using physical simulators, and pre-train CNN via transfer learning to address the scarcity of multimodal data.

Training Strategy: End-to-end joint optimization, with binary cross-entropy as the loss function. Use dropout, early stopping, and data augmentation (thermal imaging: rotation/flip/brightness adjustment; pressure: time interpolation/amplitude scaling) to prevent overfitting.

Evaluation Metrics: Focus on sensitivity (true positive rate), specificity (true negative rate), accuracy, AUC-ROC, etc., to ensure clinical reliability.

Usage Scenario: Simplified operation (start → place probe → automatic scan → display results), completed in a few minutes, operable by non-professionals, with result visualization (abnormality annotation + risk score).

Cost-Effectiveness: Hardware cost is a few hundred dollars (far lower than the tens of thousands of dollars for mammography), facilitating large-scale deployment. Early screening reduces treatment costs.

Limitations: Positioned as a screening tool; positive results require confirmation by professional examination; algorithm interpretability needs improvement.

In the future, more modalities such as micro-ultrasound probes and spectral imaging can be integrated to improve accuracy; explore advanced fusion strategies like Transformer-based multimodal encoders. This technical paradigm can be transferred to scenarios such as thyroid nodule detection and skin lesion recognition, promoting the popularization of low-cost intelligent medical devices and improving global health service accessibility.