<![CDATA[This study aims to highlight and compare the performance of three deep learning architectures, namely CNN, VGG16, and EfficientNet-B1, in classifying cataract conditions based on retinal fundus images. A total of 2600 fundus images of two classes (normal and cataract) were collected from open sources and processed in two versions: the original images and contrast-enhanced images using Contrast Limited Adaptive Histogram Equalization (CLAHE). Each model was tested using both versions of the dataset, with evaluation based on accuracy, precision, recall, and F1 score. The results of this experiment show that the application of CLAHE is proven to improve the accuracy of CNN from 0.89 (89%) to 0.97 (97%) and, importantly for clinical diagnosis, improve the recall for cataract class from 0.81 (81%) to 0.97 (97%) with precision 0.98 (98%), f1 score 0.97 (97%) and reduce the number of False Negatives (FN) from 9 to 6. Similarly, it improves the accuracy of VGG16 from 0.93 (93%) (with precision 0.91 (91%), recall 0.96 (96%), f1 score 0.94 (94%)) to 0.96 (96%) (precision 0.94 (94%), recall 0.98 (98%), f1 score 0.96 (96%), and also reduces the number of FN from 9 to 6, thereby improving clinical reliability. In contrast to the EfficientNet-B1 Model, CLAHE does not provide significant improvement. significant. significant, with an accuracy of 0.97 (97%), precision of 0.98 (98%), recall of 0.98 (98%), and f1 score of 0.97 (97%), the accuracy performance actually decreased to 0.96 (96%) and precision to 0.94 (94%). This shows that the effectiveness of preprocessing techniques is highly dependent on the model architecture used. CLAHE has been shown to be effective on conventional models such as CNN and VGG16, but is less optimal for complex pretrained models such as EfficientNet-B1. These findings contribute to the development of adaptive and efficient medical image classification systems, particularly in the context of automated cataract screening in primary healthcare.]]>
Copyrights © 2025
