Related Experiment Video
Updated: Jan 2, 2026
Discrimination and Characterization of Heterocellular Populations Using Quantitative Imaging Techniques
Published on: June 30, 2017
Comparing convolutional neural networks and preprocessing techniques for HEp-2 cell classification in
Larissa Ferreira Rodrigues1, Murilo Coelho Naldi2, João Fernando Mari3
1Departamento de Informática, Universidade Federal de Viçosa (UFV), Viçosa, MG, Brazil; Instituto de Ciências Exatas e Tecnológicas, Universidade Federal de Viçosa (UFV), Rio Paranaíba, MG, Brazil.
This study optimized deep learning models for identifying anti-nuclear antibodies in HEp-2 cells, crucial for diagnosing autoimmune diseases. Training Inception-V3 without preprocessing but with data augmentation yielded the highest accuracy (98.28%).
Area of Science:
- Immunology
- Medical Diagnostics
- Computational Biology
Background:
- Autoimmune diseases are a leading cause of mortality globally.
- Anti-nuclear antibody (ANA) detection in HEp-2 cells is a standard diagnostic method.
- Accurate classification of HEp-2 cell images is vital for reliable diagnosis.
Purpose of the Study:
- To evaluate the performance of various preprocessing strategies and convolutional neural network (CNN) architectures for HEp-2 cell classification.
- To assess the impact of enhancement methods like hyperparameter optimization, data augmentation, and fine-tuning on classification accuracy and training efficiency.
- To identify optimal deep learning approaches for automated ANA detection in HEp-2 cell immunofluorescence images.
Main Methods:
- Six image preprocessing strategies were tested.
- Five state-of-the-art CNN architectures were evaluated.
- Enhancement techniques including hyperparameter optimization, data augmentation, and fine-tuning were applied.
- Experiments were validated using a five-fold cross-validation procedure.
Main Results:
- The Inception-V3 model, trained from scratch without preprocessing and utilizing data augmentation, achieved the highest accuracy of 98.28%.
- Most CNNs performed better on non-preprocessed images when trained from scratch.
- Data augmentation consistently improved results across all evaluated models.
- Fine-tuning training reduced training time but did not enhance accuracy compared to training from scratch.
Conclusions:
- Deep learning models, particularly CNNs, show high potential for accurate HEp-2 cell classification in ANA detection.
- Training CNNs from scratch on non-preprocessed images, combined with data augmentation, is an effective strategy for maximizing accuracy.
- Data augmentation is a valuable technique for improving the performance of HEp-2 cell classification models.
- While fine-tuning can expedite training, it does not necessarily lead to superior accuracy compared to training from scratch.
