DyslexiaLens: A Multimodal Deep Learning-Powered Dyslexia Detection System

This article introduces DyslexiaLens, a production-grade backend system built with FastAPI and Docker, which uses a multimodal late-fusion CNN architecture to achieve dyslexia detection and severity scoring. The project is maintained by the DyslexiaLens team, with source code hosted on GitHub (link: https://github.com/DyslexiaLens/DyslexiaLens_AI), and was released on June 5, 2026. This system aims to lower the barrier to dyslexia screening through technical means and provide an innovative automated detection solution.

Dyslexia is a common learning disorder affecting approximately 5-10% of the global population. Traditional diagnosis relies on assessments by professional educational psychologists, which is time-consuming and costly, leading to many patients not receiving timely support. The DyslexiaLens project aims to address this issue using multimodal machine learning methods combined with visual cognitive tests and generative AI technologies.

System Architecture: Uses FastAPI (an efficient asynchronous web framework supporting inference request handling and automatic API documentation) and Docker containerization (environmental consistency, simplified deployment, horizontal scaling), integrating convolutional neural networks, traditional computer vision (CV), OCR, and generative AI subsystems.

Core Technology: Multimodal late-fusion CNN architecture. Multimodal inputs include eye-tracking, reading speed, text comprehension, and other data; late fusion allows each modality to be optimized independently before fusing in the high-level semantic space (advantages: noise resistance, learning modality interactions, supporting inference with missing modalities); CNN is used to capture local patterns and spatial hierarchical features from visually related cognitive tests.

Image Processing and CV: Processes custom grid image tests, extracts response patterns (paths, fixation points, response times), integrates OCR to convert images into structured data, combining traditional CV and deep learning.

Generative AI Application: Automatically generates test sentences (standardized materials, personalized adaptation to age/language, multilingual support).

Severity Scoring: Outputs quantitative scores (facilitating tracking), risk stratification (guiding intervention resource allocation), and dynamic assessment (analyzing change trends and intervention effects).

Deployment Considerations: Optimize inference performance (model quantization, GPU acceleration, etc.), data security (end-to-end encryption, access control), model version management (iterative updates, A/B testing), fault tolerance and degradation (maintaining core functions when subsystems are unavailable).

Application Scenarios: Large-scale school screenings (reducing costs), early intervention (gaining time), auxiliary professional diagnosis (improving efficiency, not a replacement), research tools (standardized data collection).

Note the following: 1. The model has errors (false positives/negatives; confidence levels should be labeled, and final diagnosis depends on professionals); 2. Training data may have biases, requiring fairness evaluation; 3. Cognitive data is sensitive and must comply with privacy regulations; 4. Technology should enhance rather than replace professional judgment.

DyslexiaLens demonstrates the application potential of multimodal ML in the healthcare field, integrating CV, NLP, and generative AI to provide a feasible screening solution. FastAPI and Docker ensure production readiness. For developers, it is a reference case for multi-technical stacks; for education/medical practitioners, it highlights the possibilities and boundaries of technical assistance. A balance must be struck between efficiency and diagnostic quality, as well as privacy protection.