TL;DR
This paper introduces a novel, non-invasive system for detecting brain diseases (BD) by analyzing facial color features. The system uses a specialized sensor to capture facial images, extracts color information from specific facial regions, and employs a Probabilistic Collaborative Representation Classifier (ProCRC) to distinguish between healthy individuals and those with BD, achieving high accuracy with a processing time under one minute.
What Problem Does This Paper Address?
Brain diseases (BD) are a significant global health concern, and current diagnostic methods like CT scans and MRIs are often time-consuming, inconvenient, and can be invasive or uncomfortable for patients. The authors highlight the need for a non-invasive, efficient, and patient-friendly system to detect BD. Specifically, they focus on Cerebral Infarction (stroke) as a primary target, noting that abnormal blood and oxygen supply to the brain can alter facial color and texture, providing a potential avenue for non-invasive diagnosis.
Methods
The proposed Noninvasive Brain Disease Detection System (NBDS) comprises four main components:
- Facial Image Sensor: A custom-designed 3-CCD video camera system with specialized lighting is used to capture high-quality facial images under standardized conditions. This sensor aims to minimize variations due to illumination and device dependency.
- Facial Key Block Extraction: The system automatically identifies four key facial regions (forehead, left cheek, right cheek, and nose bridge) based on pupil detection. From these regions, fixed-size "key blocks" are extracted. The paper notes that for experimental purposes, only three blocks (forehead, left cheek, and nose bridge) were ultimately used due to the symmetry of the cheeks.
- Color Feature Extraction: Color features are extracted from each facial key block. The RGB color space is converted to CIELAB and then to CIEXYZ. A color gamut consisting of six main color centroids is used to represent the color features, resulting in a six-dimensional feature vector for each block. These vectors are concatenated to form a larger feature vector.
- Classification: The Probabilistic Collaborative Representation Classifier (ProCRC) is employed for classification. ProCRC is a modification of Collaborative Representation based Classifier (CRC) that incorporates probabilistic theory, aiming for improved accuracy and speed. The system is trained and tested using this classifier.
Datasets
The study utilized a newly collected dataset comprising:
- 119 Brain Disease (BD) patients (80 diagnosed with Cerebral Infarction (CI) and 39 with other BD (OBD)).
- 595 Healthy (H) individuals.
To address the data imbalance, the healthy dataset was randomly split into five equal parts, and each part was combined with the BD dataset for training and testing. This ensured that the training and testing sets were balanced for each of the five experimental runs.
Key Insights and Ideas
- Facial Color as a Biomarker: The core idea that subtle changes in facial color, influenced by altered blood flow due to brain conditions, can be a reliable indicator for non-invasive diagnosis is a significant insight.
- Importance of Specific Facial Regions: The research demonstrates that not all facial regions are equally informative. The Left Cheek Block (LCB) emerged as the most effective single block for distinguishing BD from healthy individuals, highlighting the importance of targeted feature extraction.
- ProCRC's Effectiveness: The Probabilistic Collaborative Representation Classifier (ProCRC) proved to be a strong choice, offering a good balance between accuracy and computational efficiency compared to traditional and other representation-based classifiers.
- Trade-off between Speed and Accuracy: While some block combinations (like NBB) offered faster processing, the LCB block, despite being slightly slower, provided a substantial increase in accuracy, demonstrating a crucial trade-off that guided the selection of the optimal configuration.
Key Results
The proposed NBDS achieved impressive performance metrics:
- Overall Accuracy: 95%
- Sensitivity: 94.33%
- Specificity: 95.67%
- Average Processing Time: Less than 1 minute per sample (from image capture to detection).
The Left Cheek Block (LCB) was identified as the optimal facial key block, achieving the highest accuracy of 95% with a sensitivity of 94.33% and a specificity of 95.67% when used with ProCRC. The system also demonstrated strong performance in distinguishing sub-classes of BD:
- CI vs. H: Accuracy of 92.32%
- OBD vs. H: Accuracy of 95.88%
In classifier comparisons, ProCRC consistently outperformed k-NN, SVM, SRC, and CRC across most block combinations, particularly with LCB.
Limitations and Future Work
- Dataset Specificity: The NBDS was tested on a dataset consisting of only Chinese individuals. While the authors believe the underlying technologies are invariant to race, further validation on diverse populations is needed.
- Dataset Size: The authors plan to enlarge the dataset to include more samples for future work.
- Feature and Classifier Exploration: The paper suggests exploring more feature extractors and classifiers to further enhance BD detection.
- Integration of Other Modalities: Future work includes analyzing tongue images (also captured by the proposed device) and combining them with facial color features for more specific BD detection.
- Inclusion of Additional Health Data: The authors intend to record and analyze other individual information (physical condition, stress levels) to see if they play a role in distinguishing health status.
Summary updated