AI-Enhanced Kinetic Lateral Flow Assay (LFA) Platform

Overview

This repository hosts the deep learning framework for a next-generation Digital Point-of-Care Testing (POCT) system. Unlike traditional LFA readers that rely on static end-point imaging, this system utilizes a Video Vision Transformer (ViViT) to analyze the full kinetic reaction profile of the assay.

By capturing temporal features of the fluid dynamics, our model achieves superior sensitivity and quantification precision compared to standard colorimetric methods, specifically optimized for low-cost hardware (Raspberry Pi integration).

Key Features

• Temporal Analysis: Processes 32-frame video sequences to capture reaction kinetics using ViViT (B-16x2).
• Robust Data Loading: Custom VideoClassificationDataset handling temporal sampling, prefix alignment, and corrupted frame fallback.
• High Precision: Achieved 100% Accuracy (on validation set) for distinguishing critical concentration cutoffs.
• Hardware Optimized: Designed to work with low-resolution inputs from embedded camera modules (Raspberry Pi HQ Camera).

🧠 Model Architecture

Figure 1: High-level architecture of the proposed system. We utilize a ResNet-18 backbone for spatial feature extraction on each frame, followed by a Transformer Encoder to capture temporal dependencies across the reaction timeline.

Project Structure

• dataset.py: Custom PyTorch Dataset class with OpenCV-based video processing and temporal sampling logic.
• train.py: Training pipeline using Hugging Face Transformers and PyTorch Lightning-style loops.
• inference.py: Deployment script for batch inference on new samples.

Contributors

• Minhao Liu (Project Lead): Experimental design, hardware integration (Raspberry Pi/Microfluidics), T/C ratio algorithm, and system validation.
• Pu Sun (Model Architect): Implementation of the ViViT architecture, core network design, and customized attention mechanisms.
• Zhugang Liu (Optimization Specialist): Orchestrated the model training lifecycle, conducting extensive hyperparameter tuning and ablation studies to ensure optimal convergence and stability on the validation set.

Installation

pip install -r requirements.txt

Usage

To train the model on a new dataset:

python train.py

To run inference using trained weights:

python inference.py --weights checkpoints/best_model.pth --test_dir ./data/test