Evaluating Denoised OCT Images for AI in Diabetic Macular Edema Detection

Abstract

Purpose : This study employs artificial intelligence (AI) to distinguish between Optical Coherence Tomography (OCT) scans of individuals with Diabetic Macular Edema (DME) and those without, utilizing denoised OCT images. The research assesses whether the AI model's performance, measured through average precision, surpasses or lags behind conventional OCT images in the existing literature that investigated DME versus normal cases.

Methods : The AI model was trained using Google's collaboration platform, leveraging a publicly available image dataset from Kaggle for training. Denoised OCT images were employed to enhance accuracy. The training process, conducted on Google's servers, spanned a duration of 2 hours, with the added benefit of being both cost-free and carbon-neutral. A confusion matrix was developed and then specific metrics were calculated from the results to evaluate the model's performance.

Results : The AI model demonstrated a notable accuracy of 96%, precision of approximately 96.94%, and a sensitivity of 95%, showcasing its efficacy in distinguishing between DME and non-DME OCT scans. The model's specificity of 97% and AUC of 0.993 underscore its ability to precisely identify true negatives and exhibit robust overall discriminative power. When compared to a previous study on normal OCT images by Manikandan et al. in 2023, which yielded a lower AUC of 0.9436, our model exhibits superior diagnostic capability.

Conclusions : This study signifies a significant advancement in the application of AI for DME diagnosis through OCT scans. The utilization of denoised images, coupled with exceptional metrics, positions this model as a promising tool for enhancing the accuracy of DME detection. The comparison to a previous study highlights the improved AUC of our model, emphasizing the potential impact of using denoised OCT scans to refine diagnostic AI models in ophthalmology. The cost-free and carbon-neutral nature of the training process on Google's collaboration platform further underscores the potential of AI in resource-efficient medical applications.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Figure 1. Confusion Matrix

 

Figure 2. Model Classification of Retinal Pathologies