Improving Classification and Segmentation of Choroidal Lesions by Addressing Data Limitations with Patch-Based Approaches

Date
2024-09-19
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Choroidal nevi are benign ocular lesions that can progress to malignant forms like choroidal melanoma. Recent advancements in Deep Learning (DL) have shown potential in detecting ocular diseases, automating eye reviews, and facilitating timely treatment. However, these models require extensive labelled data, which is challenging to acquire due to the associated labeling costs. This limitation affects the optimal performance of DL models, leading to underexplored applications in this domain, with only a few studies available. This thesis presents three studies aimed at overcoming data limitations and enhancing model performance for the classification and segmentation of melanocytic choroidal tumors in fundus images. The first study involved binary classification of choroidal nevi and healthy subjects, using fundus images from the Alberta Ocular Brachytherapy dataset. Pre-trained models—ResNet50, DenseNet121, EfficientNetB7, and YOLOv8n—were evaluated. Results indicated that training on image patches rather than full-size images, combined with data augmentation to address noisy images and low-contrast lesions, resulted in the YOLOv8n model achieving the highest accuracy of 92.61%. The second study aimed to segment choroidal nevi lesions from fundus images. U-Net and YOLOv8n segmentation models were trained on the Alberta Ocular Brachytherapy dataset and validated on the Wills Eye Hospital dataset. The YOLOv8n model achieved Dice Coefficient scores of 0.833 and 0.764 for the Alberta and Wills datasets, respectively, when trained on both full-size images and image patches, along with the proposed post-processing methods. In the final study, a YOLOv8n model classified choroidal melanoma and nevi lesions from montage fundus images. The model’s performance was evaluated under two pre-training scenarios. The model pre-trained on ImageNet and fine-tuned on our dataset achieved an accuracy of 92.25%, outperforming the model pre-trained on ImageNet and fine-tuned on the Kaggle EyePACs dataset before the final fine-tuning on our dataset. Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to enhance model interpretability, providing ocular oncologists with insights into the model’s predictions. Overall, these studies emphasize the significance of DL models in automating the detection of melanocytic choroidal lesions and improving their classification and segmentation performance.
Description
Keywords
Deep Learning, Medical Image Analysis, Classification, Segmentation, Fundus Imaging, Choroidal Nevi, Choroidal Melanoma
Citation
Biglarbeiki, M. (2024). Improving classification and segmentation of choroidal lesions by addressing data limitations with patch-based approaches (Master's thesis, University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca.