THORACIS AI
The Problem: Barriers to Early Lung Cancer Detection
Lung cancer remains the leading cause of cancer death worldwide, claiming nearly 1.8 million lives annually. The stark reality is that late-stage diagnosis is the primary culprit—when caught early, survival rates can exceed 80%, but most patients are diagnosed at advanced stages where treatment options are limited.
THORACIS AI was born to address these gaps—to create a low-cost, portable, radiation‑free screening system that could be deployed anywhere, by anyone, as a first line of defense.
The Approach: Multi-Modal Fusion
Microwave Imaging
A structural sensor that detects dielectric property differences between healthy and malignant tissue (non‑ionizing, safe for repeated use)
Acoustic Analysis
A functional sensor that classifies respiratory patterns (wheezes, crackles) into disease categories
Instead of relying on a single sensing method, we took inspiration from clinical practice: doctors don't use just one tool. They combine imaging (CT, X‑ray) with auscultation (listening to breath sounds) to build a complete picture.
THORACIS AI replicates this dual‑modality reasoning through:
Multi‑Modal Fusion
A machine learning model that learns the cross‑modal relationships between structural anomalies and their acoustic signatures
The result? A screening system that achieves 99.3% accuracy in detecting tumor phantoms, a dramatic improvement over either modality alone.
Display
THORACIS AI
A look at the design and structure of THORACIS AI, combining acoustic analysis and microwave imaging for early lung health detection.
The Hardware: Built for the Real World
Every component was selected for affordability, accessibility, and performance.
Microwave Subsystem
-
4‑antenna switched array – enables multi‑angle transmission measurements
-
2–3 GHz frequency range – optimized for lung tissue penetration while staying within safe, non‑ionizing limits
-
NanoVNA-F V2 – compact, affordable vector network analyzer for S21 measurements
-
RF switches (Mini‑Circuits ZFSWA‑2‑46) – controlled by Raspberry Pi GPIO to cycle through all four antenna paths
-
Tissue‑mimicking phantoms – agar‑based materials with dielectric properties tuned to lung tissue; tumor‑mimicking inclusions with elevated water content (εr ≈ 55–60 vs. healthy εr ≈ 45–50)
Acoustic Subsystem
-
Dual BOYA BY‑M1S lavalier microphones – capture stereo lung sounds
-
Modified Primacare stethoscopes – provide professional acoustic coupling
-
USB‑C audio adapters – interface with Raspberry Pi
-
Raspberry Pi 4 – central processor and control unit
Total system cost: < $600 – a fraction of traditional alternatives, making it viable for community clinics, educational settings, and low‑resource regions.
The AI Pipeline: From Raw Data to Diagnosis
Step 1: Microwave Feature Extraction
Raw S21 data consists of 4 antenna paths × 201 frequency points = 804 values per scan. But tumors don't just affect magnitude—they also create time‑domain reflections. To capture this, we:
-
Apply Inverse Fast Fourier Transform (IFFT) to each path to convert frequency response to time‑domain signal
-
Extract 9 statistical features per path: peak amplitude, peak location, mean, standard deviation, 90th/10th percentiles, total energy, range, and squared energy
-
Concatenate the 804 frequency points with 36 time‑domain features → 840‑dimension feature vector
Background subtraction further isolates the phantom's effect by removing the baseline (air) measurement in linear scale:
S21_corrected = 10·log₁₀(10^(S21_phantom/10) − 10^(S21_air/10))
This eliminates system noise and antenna coupling, leaving only the tissue contribution.
Step 2: Acoustic Feature Extraction
Raw lung sounds are recorded at 16 kHz. We use YAMNet, a pre‑trained deep neural network, to extract 1024‑dimensional embeddings that capture subtle respiratory patterns—wheezes, crackles, and breath sounds—without requiring hand‑crafted features.
These embeddings are then fed into a small neural network (512→256→128→5 neurons) that outputs probabilities for five classes: asthma, COPD, pneumonia, healthy, and bronchial.
Step 3: Multi‑Modal Fusion
The microwave feature vector (840‑dim) and acoustic class probabilities (5‑dim) are concatenated into a single 845‑dimension fusion vector:
F = [M₈₄₀, A₅]
This combined representation is fed into an XGBoost classifier, which learns the cross‑modal relationships. For example, the model discovers that a tumor often produces acoustic patterns similar to pneumonia—likely due to airway obstruction caused by the mass. This synergy allows the system to both detect the anomaly (microwave) and classify its functional impact (acoustic).
The Results: Validation and Performance
We validated our models using rigorous cross-validation techniques to ensure reliable and generalizable results:
-
Group K-Fold / Leave-One-Group-Out
Ensures that data from the same experiment never appears in both training and validation sets, preventing data leakage and measuring true generalization -
Stratified Splits
Maintains class balance across all folds -
Class Weighting
Addresses imbalance in the acoustic dataset, improving performance on minority classes
The fusion model's 99.3% accuracy on microwave tissue classification demonstrates that combining structural and functional data yields far more reliable screening than either modality alone.
When a tumor grows, it can:
-
Obstruct airways → produce wheezing (asthma‑like pattern)
-
Cause fluid accumulation → produce crackles (pneumonia‑like pattern)
-
Irritate airways → produce chronic cough (COPD‑like pattern)
THORACIS AI learns these clinical correlations, allowing it to output a unified diagnostic suggestion: "Tumor suspected with pneumonia‑like acoustic signature"—information that a clinician can use to guide further investigation.
Our Vision
We believe that early detection shouldn't be a luxury reserved for those near major medical centers. By combining affordable hardware with intelligent software, PULMO-AI aims to:
-
Democratize screening – bring lung health assessment to underserved communities
-
Empower non‑specialists – provide decision support for community health workers
-
Reduce barriers – eliminate radiation concerns and high capital costs
-
Enable repeated monitoring – safe for regular use in high‑risk populations
Together, we can make early lung cancer detection accessible to all.
Acknowledgments
This project was developed as a research‑driven biomedical engineering initiative, combining insights from microwave engineering, acoustic analysis, and machine learning. Special thanks to mentors, open‑source contributors, and dataset providers who made this work possible.
Let’s Collaborate
If you're interested in the technical details, want to adapt the system for your own community, or have ideas for collaboration, reach out.
📧 join.detected@gmail.com
📍 Calgary, AB
📁 https://github.com/HeavenlyCloudz/THORACIS-AI