Fusing microwave imaging with acoustic analysis.
Our system doesn't just detect lung abnormalities—it combines structural and functional insights. By fusing microwave imaging (for tissue characterization) with acoustic analysis (for respiratory pattern recognition), THORACIS AI provides a comprehensive lung screening experience that mirrors how clinicians use multiple diagnostic tools.
What we make with multi-modal fusion + AI
Microwave Imaging
Uses ultra‑wideband signals (2–3 GHz) to detect dielectric property variations between healthy and malignant tissue. A 4‑antenna switched array captures S21 transmission data across 201 frequency points, and a dedicated feature extraction pipeline converts raw S21 traces into ML‑ready features (804 frequency points + 36 time‑domain features via IFFT).
Multi-Modal Fusion
The system concatenates microwave features (840‑dim) with acoustic class probabilities (5‑dim) and feeds them into an XGBoost classifier. This feature‑level fusion allows the model to learn cross-modal correlations—for example, that a tumor may produce acoustic signatures similar to pneumonia due to airway obstruction—achieving significantly higher accuracy than either modality alone.
Acoustic Analysis
A YAMNet-based deep learning model extracts 1024‑dimensional embeddings from lung sounds recorded through a modified stethoscope with dual microphones. These embeddings capture subtle respiratory patterns (wheezes, crackles) and are classified into five categories: asthma, COPD, pneumonia, healthy, and bronchial.
Interpretability & Education
The system provides confidence scores for each prediction and can generate basic 2D heatmaps indicating the most likely tumor location based on path‑wise S21 variations.
How we build reliable screening systems
Hardware validation, baseline subtraction, feature extraction, multi‑modal fusion, confidence scoring, local deployment.
Calibrate → Scan → Extract → Fuse → Interpret → Deliver
Proof in Numbers
99.3%
Fusion Accuracy
2–3 GHz
Frequency Range (Non-Ionizing)
+63.7%
Improvement over Microwave‑Only
86.4%
Acoustic Model Accuracy
<$600
Total System Cost
