A Fairness-Aware Deep Learning Framework for Multimodal Retinal Disease Classification (AMD, DR) Using OCTA Imaging

[Virtual Presenter] The development of fairness-aware diagnostic models for ocular diseases has been a significant challenge due to the lack of diverse data sets used in training these models. The absence of representative samples from various ethnic groups can lead to biased results that may not accurately reflect the actual disease prevalence among different populations. This issue is further complicated by the fact that many ocular diseases are rare and affect specific subpopulations, making it even harder to develop models that are fair and accurate..

[Audio] The use of artificial intelligence in medical imaging has been increasing rapidly over the past few years. This trend is expected to continue as AI technology advances and becomes more accessible. One area where AI is being particularly effective is in the diagnosis of age-related macular degeneration and diabetic retinopathy. These conditions are leading causes of permanent vision loss worldwide. Optical coherence tomography angiography (OCTA) is a powerful tool that provides detailed images of the retina and choroid. However, despite its benefits, AI models trained on OCTA images have shown inconsistent results when applied to different demographics. Several factors contribute to these inconsistencies, including biased training data, variations in image acquisition, and differences in disease prevalence among population groups. These biases can further disadvantage vulnerable populations, who already bear a greater burden of these conditions. The lack of standardization in AI model development and deployment can exacerbate these issues. Furthermore, the need for diverse and representative training datasets is critical to ensure accurate diagnoses..

[Audio] The primary objective of this project is to develop a multimodal deep learning framework for classifying AMD severity and DR stage using OCTA-derived vascular and structural biomarkers. This framework should incorporate explicit fairness constraints to ensure equal treatment for all individuals, regardless of their demographic characteristics. The secondary objectives of this project are to quantify and mitigate performance disparities across different demographic and clinical subgroups. Furthermore, the project aims to introduce interpretability methods to identify fairness-critical features, which will enable better decision-making and more accurate predictions. Another key aspect of this project is to validate the generalizability of the proposed framework across multiple acquisition sites and OCTA devices. This will help to ensure that the framework can be applied consistently and accurately across different settings..

[Audio] The data curation and harmonization process involves aggregating multiple datasets from various centers and augmenting them with relevant demographic information. This process aims to minimize biases associated with specific scanners and sites by applying domain adaptation techniques such as CycleGAN and AdaIN. The approach also incorporates fairness-aware deep learning architecture, which utilizes 3D CNNs or ViTs to analyze facial and cross-sectional OCTA images. Fairness is integrated into the methodology through several steps including reweighting and resampling sensitive attributes, constrained optimization to ensure demographic parity, and threshold tuning to achieve equalized predictive parity across subgroups..

[Audio] The AI system uses a combination of fairness metrics and secondary metrics to evaluate its performance. The primary metrics include measures such as equal opportunity difference, average odds difference, and disparate impact ratio. These metrics help identify biases in the data and ensure that the model is fair and unbiased. Secondary metrics focus on stratified performance across subgroups, calibration curves, interpretability, and clinical validation. The AI system also employs SHAP values and attention mapping to pinpoint fairness-sensitive regions. Furthermore, clinician-in-the-loop validation is used to ensure that the model is clinically plausible and effective in mitigating bias..

[Audio] The project aims to address issues related to bias in ophthalmic AI models. The team has identified several key areas where bias can occur, including demographic characteristics, medical conditions, and treatment options. To mitigate these biases, the team plans to develop a framework for detecting and addressing bias in AI models. This framework will include tools for data preprocessing, feature engineering, and model evaluation. The team also intends to collaborate with experts from various fields, including medicine, computer science, and social sciences, to ensure that the developed solutions are effective and practical. Furthermore, the team plans to conduct thorough testing and validation of the proposed solutions to ensure their accuracy and reliability..

[Audio] The project aims to address the issue of unequal access to eye care services among different populations. The initiative focuses on developing an AI-powered diagnostic tool that can detect diseases such as glaucoma and diabetic retinopathy. The tool will be designed to ensure fairness in its decision-making process, taking into account factors such as age, sex, and socioeconomic status. The goal is to create a more equitable distribution of eye care resources, reducing the burden on already disadvantaged groups. The project also incorporates principles of fairness by design, which involves embedding ethical considerations into the development process from the outset. This approach enables the creation of AI systems that are transparent, accountable, and fair. The project's alignment with key initiatives such as the NIH's UNITE Initiative and the WHO's Global Action Plan for Universal Eye Health underscores its importance. By addressing the issue of unequal access to eye care, the project has the potential to improve the lives of millions of people worldwide..