Page 43 - AIH-1-4

Artificial Intelligence in Health Segmentation and classification of DR using CNN

images and their corresponding ground truth masks. Table 1. Segmentation results
Utilizing the U-Net model previously trained on augmented
data, the model’s predictive capabilities are scrutinized for Segmentation Metric(s) Performance
method
pixel-level segmentation accuracy. The model’s state is
restored using the best-performing checkpoint achieved UNet Dice coefficient; IoU 0.95 dice; 0.92 IoU
during the training phase, ensuring the evaluation is based DeepLab Dice coefficient; IoU 0.80 mIoU
on the most optimized configuration. FCN Pixel accuracy; IoU 0.92-pixel accuracy; 0.70 IoU
SegNet Pixel accuracy; IoU 0.88-pixel accuracy; 0.65 IoU
For each test sample, the retinal fundus image is pre-
processed by normalizing pixel values and transposing the PSPNet Mean intersection 0.78 mIoU
over union (mIoU)
channels to match the model’s input requirements. Similarly,
the ground truth mask undergoes pre-processing to LinkNet Pixel accuracy; IoU 0.85-pixel accuracy; 0.68 IoU
facilitate direct comparison with the model predictions. The Mask R-CNN AP 0.65 AP
evaluation metrics, including Jaccard index, F1 Score, recall, Abbreviations: IoU: Intersection over union; mIoU: Mean intersection
precision, and accuracy, are computed for each test image. over union; AP: Average precision; FCN: Fully Convolutional Network.
These metrics quantify the model’s ability to accurately
delineate DR-related regions in the retinal fundus images. 35 Table 2. Performance metrics for lesion segmentation in the
Indian diabetic retinopathy image dataset
The computational efficiency of the model is also
assessed through the calculation of frames per second Different Jaccard F1 score Recall Precision Accuracy
(FPS) during the inference process. This metric provides segmentations score score
insights into the real-time processing capabilities of Blood vessel 0.6634 0.7974 0.7771 0.8240 0.9922
the model, offering valuable information for potential Hard exudate 0.6663 0.7953 0.7634 0.8252 0.9986
deployment in clinical or real-world scenarios. Visual Soft exudate 0.6679 0.7874 0.7738 0.8340 0.9981
representations of the model’s predictions, alongside the Hemorrhage 0.6551 0.7874 0.7852 0.8161 0.9958
original retinal fundus images and ground truth masks, are Microaneurysms 0.6761 0.8061 0.7731 0.8279 0.9967
saved for qualitative analysis and comparison. Optical disc 0.6638 0.7956 0.7671 0.8238 0.9989

3. Results
3.1. Segmentation results Notably, the model demonstrated efficient segmentation
with an average accuracy of 0.9986 and an impressive
UNet achieved the highest performance with a dice FPS rate of 361.188719052. Visual results were generated
coefficient of 0.95 and an intersection over union (IoU) for each test image, illustrating the original image, the
of 0.92. Following closely, fully convolutional network ground truth mask, and the predicted segmentation
exhibited a pixel accuracy of 0.92 and an IoU of 0.70. mask. Overall, these findings highlight the effectiveness
DeepLab attained a mean IoU (mIoU) of 0.80, while and computational efficiency of the proposed blood
SegNet and LinkNet demonstrated pixel accuracies of 0.88 vessel segmentation model, showcasing its potential for
and 0.85, respectively, with IoUs of 0.65 and 0.68. PSPNet applications in DR diagnosis and treatment planning, as
yielded a mIoU of 0.78. Finally, Mask R-CNN, evaluated visualized in Figure 6.
by average precision, achieved a performance of 0.65. In this research, we conducted a comprehensive
These results provided insights into the efficacy of different evaluation of a segmentation model across various retinal
segmentation methods, highlighting UNet as a particularly 15 14
promising approach for accurate image segmentation tasks structures, including blood vessels, EXs, SEs, HEs,
MAs, and the optical disc. The model’s performance was
37
in Tables 1 and 2.
quantitatively assessed using key metrics, revealing high
In this study, we employed a U-Net-based model segmentation accuracy across all structures. The Jaccard
for the segmentation of blood vessels in retinal images. scores ranged from 0.6551 to 0.6761, indicating substantial
The implemented model was evaluated on a test dataset overlap between the predicted and ground truth masks.
comprising retinal images and corresponding ground The model achieved notable F1 scores, demonstrating
truth masks. The testing process involved loading images a harmonious balance between precision and recall,
and masks, pre-processing the data, and utilizing a ranging from 0.7874 to 0.8061. Particularly, commendable
pre-trained U-Net model for predictions. The model’s was the recall scores, signifying the model’s ability to
performance was assessed using several metrics, including correctly identify relevant instances, with values ranging
Jaccard similarity, F1 score, recall, precision, and accuracy. from 0.7634 to 0.7852. Precision scores, representing the

Volume 1 Issue 4 (2024) 37 doi:10.36922/aih.2783