ANAS-mart (Antinuclear Antibody Pattern Classification using Artificial Intelligence)

Autoimmune diseases have an estimated prevalence between 3-5%, generating a significant impact on quality-adjusted life years (QALYs) and disability-adjusted life years (DALYs) (Palmezano et al., 2018). Given this burden, early diagnosis is essential to enable appropriate treatment and follow-up, thereby reducing the sequelae associated with these diseases. Accurate interpretation of antinuclear antibodies (ANAs) requires a rigorous reading of the patterns.
Our project aims to evaluate the performance of an ANAs pattern classifier using artificial intelligence models applied to indirect immunofluorescence microscopy (IFA) images. For this purpose, a database composed of 300 images labeled according to standard ANAs patterns and enriched with relevant clinical data will be constructed. These images will be preprocessed to ensure uniformity in size, scale and position, and key geometric features such as morphology, sizes and distances will be extracted. The images and their features will be used as inputs to train two types of classification models: multilayer neural networks (MLN) and Random Forest (RF).
The performance of the models will be evaluated through metrics such as precision, sensitivity and accuracy, considering the limitations of a reduced dataset. This project fosters interdisciplinary collaboration: medicine and engineering and represents a first step towards accessible tools for expert analysis of IFI images. The results obtained will serve as a basis for future developments, including the implementation of advanced techniques such as deep learning with larger datasets.
Sustainable Development Goal 3 (SDG3).
4.1 IFI IMAGE PROCESSING FOR ANAs
The method that was used to obtain the images already existing in the CREA information bank, and that will be used to increase the existing N up to 300 images, (in case any of the images does not meet the necessary conditions for inclusion), is IFI. This method allows screening and semi-quantitative determination of ANAs in human serum. In order for the samples to be in suitable conditions and for the performance of the test to be reliable at the time of analyzing the results, the following must be taken into account:
1.A suitable substrate (human HEp-2 epithelial cells).
The evaluation of the fluorescence/protein (F/P) ratio.
The subclass specificity of the immunoglobulin of the conjugate (Yumpu.com, n/d).
In the IFA technique, serum samples are incubated with the HEp-2 cell line, then washed to remove unbound and non-specific antibodies. Subsequently, they are incubated with a fluorochrome-labeled conjugate, then washed to remove the unbound conjugate and immunocomplex. Subsequently, the samples are observed under the IFA microscope which shows if the tests are positive when there is emission of fluorescence of an apple green color, in the areas of the cytoplasm or nucleus, giving one of the characteristic patterns.

For interpretation, two possible results are considered: a sample in which the fluorescence emission is less than or equal to that observed in the negative control is classified as negative, and a sample showing fluorescence greater than the control is classified as positive. A positive sample is classified according to the pattern and the weighting of its intensity (Yumpu.com, n.d.), as follows:
Intensity weighting: The following criteria are available
4.4+ Bright apple-green fluorescence
5.3+ light apple-green fluorescence
6.2+ Clearly positive fluorescence
7.1+ Fluorescence that clearly differentiates a nuclear and/or cytoplasmic pattern from the background structure
Patterns evidenced: The International consensus on ANA patterns (ICAP) will be used to classify the patterns. This classification has 3 general groups: nuclear, cytoplasmic and mitotic, which have codes from AC-1 to AC-29 (ANA Patterns, n.d.).

For our project, in principle, the algorithm will be trained in the identification of the main nuclear and cytoplasmic patterns.
Finally, the quality control is given by two criteria, if for any reason any of them is not fulfilled the result will be invalid and the assay must be repeated. This process is performed on each slide to ensure its quality (Yumpu.com, n.d.).
El positive control of standard ANAs, undiluted must have an intensity greater than or equal to 3+.
El negative ANA standard control should emit no fluorescence or minimal fluorescence (known as “background noise”).
The immunofluorescence images were previously acquired and the term “processing” refers to the digital treatment that is subsequently performed on these images using software such as LAS X (Leica Application Suite X) or ImageJ. This treatment includes brightness/contrast adjustments, fluorescence intensity measurement, and signal quantification, without altering the original data. Image processing is detailed below:
Image acquisition
Images were previously acquired by fluorescence inverted microscopy using a DMi8 microscope (Leica) and LAS X software. Standardized parameters of exposure, gain and laser settings were used, according to the experimental design, in order to ensure comparability between samples.
Image quality review
The quality of the raw images will be evaluated to verify the absence of artifacts, signal saturation or blurring.
Only images with adequate resolution and a clear specific signal will be included.
Format conversion
Images will be exported from LAS X (original .lif format) and converted to high resolution .TIF format, compatible with other analysis programs such as ImageJ.
General brightness and contrast adjustments
In LAS X or ImageJ, uniform brightness and contrast adjustments will be applied to all images in the same experiment for viewing purposes only.
These adjustments are made in LAS X or ImageJ.
The percentage of pixels allowed to saturate will be specified to optimize the overall contrast of the images.
Normalization will consist of recalculating the intensity values of each pixel, generating a new version of the image with comparable scales between samples.
Background reduction
Methods will be applied to eliminate the electrical noise generated by the detection system.
This stage seeks to reduce the variability associated with noise and improve the signal-to-noise ratio (SNR) of the images.
Signal quantification
Regions of interest (ROIs) will be defined manually or by automatic thresholds.
The mean fluorescence intensity of the channel within the ROIs will be measured.
The values obtained will be normalized against internal controls or the corresponding area of the ROI.
Documentation
All processing parameters will be documented in detail.
Both the original versions and the processed images will be stored separately to ensure traceability and reproducibility of the analysis.
After image processing, the interpretation of the results will be performed with the corresponding review of ANAs patterns (as mentioned in the project). The main stages of the study to develop and evaluate classifiers based on RNM and Random Forest (RF) with IFA images of patients with ANAs patterns are described below.
4.2 CONSTRUCTION OF THE IFA IFA IMAGE DATA SET
300 IFA microscopy images from patients with a confirmed diagnosis of SLE will be collected from the CREA data bank. Each image will represent a microscopic field containing specific fluorescence patterns, which will be further classified according to the main ANAs patterns defined by ICAP.
4.3. LABELING IMAGES WITH ANAs PATTERNS
Each image will be labeled by experts, manually, with one of the main existing ANAs patterns. This process will include consensus among evaluators and, in case of discrepancy, predefined classification criteria will be used. This stage is crucial to ensure the quality of the labels that will serve as the basis for the supervised training of the algorithms.
4.4. PREPROCESSING OF THE IMAGES
Preprocessing will ensure uniformity in the technical characteristics of the images. The steps include:
-Redimensioning: images will be adjusted to a standard size (256x256 pixels) to facilitate processing.
-Grayscale conversion: Color channels will be removed to work with a single intensity dimension.
-Centering: Relevant patterns will be positioned in the center of the image, eliminating unnecessary noise.
-Normalization: Pixel values will be normalized to improve convergence during algorithm training. These operations will be performed using Python libraries or Matlab image toolbox in order to automate the workflow.
4.5. RNM MODEL TRAINING
An RNM model will be implemented using Python (TensorFlow or PyTorch) or MATLAB (Deep Learning Toolbox). The architecture will include an input layer, fully connected hidden layers, and an output layer with defined nodes, each representing an ANAs pattern (Goodfellow et al. 2016).
-The model will be trained using backpropagation and optimizers such as Adam or SGD.
-Triggering functions such as ReLU in the hidden layers and softmax in the output layer will be used.
-The data set will be divided into training (70%), validation (15%) and testing (15%).
4.6. EXTRACTION OF GEOMETRIC FEATURES FROM IFI IMAGES
To enrich the data, geometric features will be extracted from the images (Gonzalez & Woods 2018; Pedregosa et al. 2011), such as:
-Morphology: shape and contour of fluorescent patterns.
-Size: Areas and perimeters of regions of interest.
-Distances: Spacings and relative distributions between fluorescent structures. These features will be obtained using tools such as scikit-image in Python or custom functions in MATLAB. This stage provides complementary descriptors for the RF model.
4.7. RF MODEL TRAINING
An RF model will be built and trained using Python (scikit-learn) or MATLAB (Statistics and Machine Learning Toolbox). The geometric features extracted in the previous stage will serve as input to the model. Model hyperparameters, such as the number of trees and maximum depth, will be optimized by cross-validation. For classification, information gain measures or Gini index for node splits will be used. (Breiman 2001; Pedregosa et al. 2011)
4.8. PREDICTION WITH THE RNM AND RF MODELS
Both trained models will be applied to the test set to generate predictions. The RNM model will directly classify the images, while the RF model will make predictions based on geometric features. The probabilities of each class will be stored along with the predicted labels.
4 9. EVALUATION OF THE MODELS
The performance of both models will be evaluated using standard metrics:
-RNM: Accuracy, precision, sensitivity, F1-score, ROC curve and confusion matrix.
-RF: The same metrics will be calculated to compare their performance based on geometric features. Evaluations will be performed using scikit-learn or MATLAB.
4 9. EVALUATION OF THE MODELS
The performance of both models will be evaluated using standard metrics:
-RNM: Accuracy, precision, sensitivity, F1-score, ROC curve and confusion matrix.
-RF: The same metrics will be calculated to compare their performance based on geometric features. Evaluations will be performed using scikit-learn or MATLAB.
4.10. MODEL COMPARISON
The results of the two models will be compared in terms of their performance metrics and their ability to correctly classify the predominant patterns in the dataset. In addition, the usefulness of geometric features as an alternative input to the RF model will be evaluated.

Autoimmune diseases
Immunology
ANAs
ANAs patterns
Artificial intelligence
AI
Indirect immunofluorescence
IFI
Multilayer neural networks
Random Forest

OBJETIVO GENERAL:

•Evaluar el desempeño de un clasificador de patrones de ANAs usando modelos de inteligencia artificial en imágenes de IFI.

OBJETIVOS ESPECÍFICOS:

•Construir una base de datos con las imágenes de microscopía IFI asociadas con patrones de ANAs y datos clínicos.
•Extraer características geométricas relevantes que describan los patrones de las imágenes IFI.
•Entrenar los algoritmos de clasificación de patrones ANAs usando modelos de RNM y RF.
•Determinar las métricas de desempeño de los modelos RNM y RF.
•Potenciar en los estudiantes habilidades transversales en el desarrollo de un proyecto de investigación.
•Fomentar la colaboración entre profesores y estudiantes en el proyecto que redunde en soluciones innovadoras que involucren la inteligencia artificial y que beneficien directamente a pacientes con enfermedades autoinmunes.
•Introducir a los estudiantes en el área de inteligencia artificial, iniciando con métodos sencillos, como redes neuronales simples, para garantizar un aprendizaje progresivo y efectivo.

Resultados esperados:
1. Base de datos de imágenes de microscopía de IFI asociada con patrones de ANAs y datos clínicos.
2. Matrices de pesos entrenadas del modelo de RNM generado para la para clasificación de patrones de ANAs.
3. Árbol de decisión entrenado del modelo RF generado para la para clasificación de patrones de ANAs.
4. Divulgación de ciencia abierta: Considerando los cinco pilares fundamentales que lo conforman, el proyecto ANAS-mart materializará el diálogo con la comunidad científica, entre profesionales clínicos, estudiantes, personal de laboratorios de investigación, empresas vinculadas y proveedores de microscopía, fomentando así la interdisciplinariedad. Esto a través de videoconferencias, congresos y charlas dirigidas a los actores previamente mencionados. En estos espacios se abordará el impacto del proyecto en el entrenamiento de una herramienta costo eficiente que mejore la clasificación de los patrones de ANAs, resaltando las métricas de nueva generación. Así como también se destacan las buenas prácticas de la integridad científica, que se manejarán durante el desarrollo del proyecto.
Por otro lado, se fortalecerán las bases de datos y los formatos donde se almacenará la información generada durante el desarrollo del proyecto, asegurando que sean accesibles, reutilizables, reproducibles y conservados a largo plazo. Este enfoque busca facilitar la reutilización equitativa de los contenidos, promoviendo la diversidad y la inclusión, que son característicos de la comunicación académica y los datos de investigación en abierto. (Universidad del Rosario. 2020)