Bias Detection in Histology Images Using Explainable AI and Image Darkness Assessment.

Area of Science:

• Medical Artificial Intelligence
• Computational Pathology
• Biomedical Imaging Analysis

Background:

• Medical AI models can exhibit biases, impacting fairness and clinical utility.
• Cervical histology image classification is susceptible to biases, including those related to image quality.
• Explainable AI (XAI) offers tools to understand and address AI model decision-making.

Purpose of the Study:

• To present a novel framework combining XAI and image darkness assessment for detecting and mitigating bias in cervical histology image classification.
• To evaluate the effectiveness of bias mitigation strategies on deep learning model performance.
• To improve the fairness, generalizability, and clinical utility of AI models in pathology.

Main Methods:

• Employed four deep learning architectures: AlexNet, ResNet-50, EfficientNet-B0, and DenseNet-121.
• Utilized Grad-CAM and saliency maps for bias identification within model predictions.
• Implemented brightness normalization and synthetic data augmentation for bias mitigation.

Main Results:

• EfficientNet-B0 achieved the highest accuracy after bias mitigation.
• Mitigation strategies successfully shifted model focus towards clinically relevant features.
• Statistical analysis (ANOVA) showed a significant reduction in image darkness influence (F-statistic decreased from 120.79 to 14.05).

Conclusions:

• The proposed framework effectively detects and mitigates bias in cervical histology image classification AI.
• Bias mitigation improves model accuracy, fairness, and alignment with clinical features.
• XAI combined with image analysis is a promising approach for developing robust and equitable medical AI.