Page 42 - AIH-1-4

Artificial Intelligence in Health Segmentation and classification of DR using CNN

Figure 4. The picture illustrates the architecture of U-Net layers used for segmentation. Image reprinted with permission, Copyright © 2015, Springer
International Publishing Switzerland. 38
Abbreviations: Conv: 3 ×x 3; ReLU: Rectified linear unit.

with their corresponding masks. To offer a comprehensive with the lowest validation loss, serves as a pivotal asset for
overview of the experimental setup, key hyperparameters future deployment and analysis. 33
and insightful statistics about the dataset are presented.
The choice of computation device, which is based on the 2.4.3. Post-training
availability of a Compute Unified Device Architecture- The post-training phase marks the culmination of the
enabled graphic processing unit, is disclosed, and the training experiment, where the training outcomes are
U-Net model is adeptly moved to the selected device. meticulously summarized, and the final model is prepared
The initialization of the Adam optimizer, the learning for deployment or further analysis. The best-performing
rate scheduler, and the combination of dice loss and model is meticulously selected based on the lowest
binary cross-entropy provide a robust foundation for the validation loss achieved during training and is safeguarded
subsequent training phases. 31 as a checkpoint for subsequent use. Key metrics, such as
training loss, validation loss, and epoch-wise training
2.4.2. Main training times, are presented to provide a holistic evaluation of the
The main training phase unfolds over a specified model’s performance. Furthermore, this phase allows
18
number of epochs, wherein the U-Net model undergoes for insights into the training process, including potential
iterative training. Within each epoch, the model is improvements or challenges faced, fostering a deeper
33
rigorously trained using the prepared training dataset, understanding of the model’s behavior in the context of DR
and the optimizer diligently works to minimize the detection (Figure 5). The post-training phase thus solidifies
loss, computed through a combination of the dice loss the experiment’s completion, with the trained U-Net
and binary cross entropy. Simultaneously, the model’s model ready for deployment in practical applications or
proficiency is rigorously evaluated on the validation further investigative studies. 34
dataset to monitor its generalization capabilities. The
18
training process is intricately monitored with detailed 2.5. Testing
output, including epoch-wise loss values and elapsed In the testing phase, the trained U-Net model is rigorously
time, fostering a nuanced understanding of the model’s evaluated on a separate dataset to assess its performance
convergence patterns. The strategic selection of the best in semantic segmentation of retinal fundus images for
model checkpoint ensures that the model attains optimal DR detection. The experiment involves loading the
performance. This checkpoint, capturing the model’s state preprocessed test dataset, consisting of retinal fundus

Volume 1 Issue 4 (2024) 36 doi:10.36922/aih.2783