Subramanya B, Shivanna Divya B, G Nithin Raj, Prabhu Pratham S, Yaseer Mohammed, Rao Roopa S
Department of Computer Science and Engineering, Faculty of Engineering & Technology, Ramaiah University of Applied Sciences, Bengaluru, India.
Department of Oral Pathology, Faculty of Dental Sciences, Ramaiah University of Applied Sciences, Bengaluru, India.
Mediterr J Rheumatol. 2025 Mar 31;36(1):50-62. doi: 10.31138/mjr.060624.doa. eCollection 2025 Mar.
The diagnosis of autoimmune disorders, particularly through the Anti-Nuclear Antibodies (ANA) Indirect Immunofluorescence (IIF) test utilising human epithelial type-2 (HEp-2) cells, presents a formidable challenge due to the subjective nature of pathologists' analysis. In response, this study proposes an innovative automated approach that integrates deep learning, advanced image processing, guided Hep-2 Cell, and mitotic cell instance segmentation.
Leveraging the ICPR 2016 dataset for training and evaluation, this research encountered an initial challenge of dataset imbalance, with a significantly lower number of mitotic cells compared to HEp-2 Homogenous cells. To overcome this, data augmentation techniques were strategically employed to ensure a balanced representation.
In Experiment 1, the Detectron2 model achieved an overall mean Average Precision of 54% for segmentation masks and 55% for bounding boxes. In Experiment 2, the YOLOv8n model demonstrated an impressive overall Mean Average Precision score of 94% for bounding boxes and 93% for segmentation masks, showcasing its exceptional efficacy in detecting HEp-2 cells and mitotic cells. The instance segmentation provided a more granular analysis, revealing the count of cells in each class, further highlighting the model's proficiency in diagnosing autoimmune diseases.
This study establishes a reliable and automated method for HEp-2 Homogenous cell detection, addressing the ongoing challenges in autoimmune disease diagnosis and contributing significantly to the ongoing revolution in this critical field.
自身免疫性疾病的诊断，尤其是通过利用人上皮2型（HEp-2）细胞的抗核抗体（ANA）间接免疫荧光（IIF）试验进行诊断，由于病理学家分析的主观性而面临巨大挑战。作为回应，本研究提出了一种创新的自动化方法，该方法集成了深度学习、先进的图像处理、引导式HEp-2细胞和有丝分裂细胞实例分割。
利用ICPR 2016数据集进行训练和评估，本研究最初遇到了数据集不平衡的挑战，与HEp-2同质细胞相比，有丝分裂细胞的数量明显较少。为克服这一问题，策略性地采用了数据增强技术以确保平衡的代表性。
在实验1中，Detectron2模型的分割掩码总体平均精度为54%，边界框为55%。在实验2中，YOLOv8n模型的边界框总体平均精度得分高达94%，分割掩码为93%，展示了其在检测HEp-2细胞和有丝分裂细胞方面的卓越功效。实例分割提供了更细致的分析，揭示了每个类别的细胞数量，进一步突出了该模型在诊断自身免疫性疾病方面的熟练程度。
本研究建立了一种可靠的自动化HEp-2同质细胞检测方法，解决了自身免疫性疾病诊断中持续存在的挑战，并为这一关键领域正在进行的变革做出了重大贡献。
