Detection of Auto-Immune Disease using Deep Learning Techniques.

Area of Science:

• Medical Diagnostics
• Computational Pathology
• Immunology

Background:

• Autoimmune disorder diagnosis relies on Anti-Nuclear Antibodies (ANA) Indirect Immunofluorescence (IIF) tests using human epithelial type-2 (HEp-2) cells.
• Pathologist subjectivity in analyzing HEp-2 cell images poses a diagnostic challenge.
• Automated methods are needed to enhance the accuracy and efficiency of autoimmune disease diagnosis.

Purpose of the Study:

• To develop and evaluate an automated deep learning approach for HEp-2 cell and mitotic cell instance segmentation.
• To improve the reliability and objectivity of autoimmune disorder diagnosis through image analysis.
• To address the challenge of dataset imbalance in automated cell detection.

Main Methods:

• Utilized the ICPR 2016 dataset for training and evaluation.
• Employed data augmentation techniques to balance the dataset, particularly for mitotic cells.
• Implemented and compared deep learning models, including Detectron2 and YOLOv8n, for instance segmentation.

Main Results:

• The YOLOv8n model achieved high performance with a Mean Average Precision (mAP) of 94% for bounding boxes and 93% for segmentation masks.
• Detectron2 showed lower performance with mAP of 54% for masks and 55% for boxes.
• Instance segmentation enabled granular cell counting, demonstrating proficiency in HEp-2 cell and mitotic cell detection.

Conclusions:

• An automated, reliable method for HEp-2 cell detection was established using deep learning.
• The developed approach significantly enhances the accuracy of autoimmune disease diagnosis.
• This work contributes to the advancement of automated diagnostic tools in clinical pathology.