Browsing by Author "Abdul Rahman Naveed"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Uncertainty information for diabetic retinopathy detection
    (UMT.Lahore, 2024) Mohsin Akram; Ahtisham Khan; Abdul Rahman Naveed
    Deep Learning (DL) has played a pivotal role in advancing medical imaging, revolutionizing disease diagnosis and treatment. In the realm of Diabetic Retinopathy (DR), DL models have exhibited remarkable efficacy in tasks like classification, segmentation, detection, and prediction. However, the inherent opacity and complexity of DL models can lead to decision-making errors, particularly in intricate cases, necessitating the estimation of uncertainty in predictions. Classical DL models alone struggle to provide this estimation adequately. To address this challenge, Bayesian DL methods have emerged and are gaining prominence in the field. This research introduces a DenseNet-121 pre-trained (CNN) model using transfer learning for DR classification, employing a simple architecture. Subsequently, Bayesian extensions are applied to this DenseNet-121 model, utilizing Monte Carlo Dropout (MC-Dropout) method. This Bayesian enhancement aims to capture the posterior predictive distribution. The evaluation is conducted on our preprocessed merged dataset (APTOS-2019 + DDR), gauging the models' performance. Experimental results unveil the superiority of the proposed models over other state-of-the-art counterparts in terms of test accuracy. Specifically, the simple DenseNet-121 model achieves an accuracy of 91.23%, BCNN MC-Dropout achieves 97.33%. Additionally, the study computes entropy and probabilities on the predictive distribution to quantitatively measure model uncertainty. This research underscores the potential advantages of integrating Bayesian DL methods into medical image analysis, offering a pathway to augment accuracy and reliability in disease diagnosis and treatment. The proposed models not only outperform existing models but also contribute to the ongoing exploration of uncertainty estimation in DL applications, crucial for enhancing clinical decision-making.