Tearz

GIF
Usage Showcase
Healthy Patient
Unhealthy Patient

Inspiration

Early detection of systemic diseases like Sklerosis Multiplex and Diabetes often requires invasive or costly tests. We realized that human tears are a "liquid biopsy" containing biochemical maps of our health. We built Tearz to turn a single drop of fluid into an instantaneous, non-invasive diagnostic window.

What it does

Tearz is a two-stage diagnostic platform:

Screener: A high-sensitivity model that flags "Unhealthy" subjects.
Specialist: A deep-learning ensemble that identifies four specific pathologies: Glaucoma, Dry Eye (SucheOko), Sclerosis, and Diabetes. It features a drag-and-drop medical GUI that provides clinicians with predictions, confidence scores, and diagnostic recommendations.

How we built it

Preprocessing: A Gwyddion-inspired pipeline (Pseudo-field flattening, Median row alignment, CLAHE, and Skeletonization) to isolate crystalline structures from sensor noise.
Feature Engineering: Extracted 240 topological and texture features (Fractal Dimension, J/E Ratio, Haralick textures) for a robust Stage 1 Logistic Regression.
Deep Learning: A 5-fold ResNet50 Ensemble using transfer learning, label smoothing, and Test-Time Augmentation (TTA) to maximize classification accuracy.

Challenges we ran into

Sensor Noise: Surgical removal of 1-pixel wide scanner "scars" without blurring delicate fern branches.
Data Scarcity: Preventing overfitting and data leakage on a small dataset (169 images) by enforcing strict subject-level splits.
Computation Time: Training on raw 16-bit images was limited by available hardware, necessitating a strategic pivot to an optimized 8-bit pipeline.

Accomplishments that we're proud of

Achieving a perfect safety net for initial pathology screening.

Resilient F1-Score: Reaching a 72.81% F1-score on a multi-class problem with limited subject diversity.
Clinical Utility: Bridging the gap from a Jupyter notebook to a functional, user-friendly medical interface.

What we learned

Data Quality > Model Complexity: Surgical preprocessing and row alignment were more impactful than switching to larger Transformer models.
Small Data Strategy: Handcrafted topological features provide a stronger inductive bias than raw pixels in low-sample medical regimes.
Medical Logic: The importance of designing for "Explainability" and acknowledging "Confounders" like age and hydration.

What's next for Tearz

Dataset & Physiological Expansion: Scaling subject diversity and integrating patient metadata (age, BMI, hydration) to better "de-confound" metabolic shifts from specific disease signatures.
16-Bit High-Fidelity Training: Utilizing upgraded computational power to train on raw 16-bit data, capturing 65,536 intensity levels to reveal nearly invisible crystalline markers.
Advanced Architecture Fine-Tuning: Leveraging high-performance computing to explore larger Vision Transformers and ensembles, pushing our F1-scores toward clinical-grade 90%+.