This section investigates the significance of our findings, examine their implications for glaucoma diagnosis, and offers critical insights into the performance and potential limitations of our methodology. It validates the significance of our findings and lays the door for further development of automated methods for glaucoma diagnosis and optic nerve segmentation.
4.1 Hyper-parameters Details
The hyper-parameters [35] utilized by the combination UNet-Inception architecture-based technique for intraocular fragmentation with Harris Hawk selecting features are listed in Table 2. The learning rate, batch size, optimization, number of epochs, patch size, and activation function are some of the model parameters.
Table 2: Hyperparameters details
Hyper parameter
|
Value
|
Hyper parameter
|
Value
|
Input image size
|
512x512
|
Dropout rate
|
0.2
|
Batch size
|
8
|
Inception block filter sizes
|
[3x3, 5x5, 7x7]
|
encoder blocks
|
4
|
filters in Inception block
|
16
|
filters in first encoder block
|
32
|
Use batch normalization
|
Yes
|
filters in last encoder block
|
512
|
Use skip connections
|
Yes
|
decoder blocks
|
4
|
Use attention layer
|
Yes
|
learning rate
|
0.001
|
Optimizer
|
Adam
|
Learning rate schedule
|
ReduceLROnPlateau
|
Loss function
|
0.9*Dice Loss + 0.1*Focal Loss
|
While the batch size dictates the quantity of specimens utilised in each iteration process, its training error controls the optimizer's step length throughout training. Adam, a popular optimization in transfer learning, was the engine employed in this work. The quantity of epochs indicates how frequently the complete training dataset is used. The dimensions of the input picture chunks for use during learning, sometimes referred to as the "patch size," has an effect on the compromise between speed and segmentation accuracy.
4.2 Evaluation Parameters
The significance of Accuracy, Dice Score, and Sensitivity in the context of Glaucoma Segmentation using the proposed Novel UNet-Inception Attention Architecture can be explained as follows:
i. Accuracy: metric used to evaluate the overall correctness of a segmentation model. It measures the ratio of correctly predicted pixels (both true positives and true negatives) to the total number of pixels in the image as given in Eq. (1) [35]. In Glaucoma Segmentation, accuracy indicates how well the model is able to correctly identify both the regions with and without glaucoma in the retinal images. A high accuracy score implies that the model is effectively distinguishing between healthy and glaucomatous areas.
ii. Dice Score (F1 Score): also known as the F1 Score is a metric that balances both precision and recall [36]. It is calculated as 2 times the intersection of the predicted and true positive regions divided by the sum of the predicted and true positive regions. The Dice Score is crucial as it measures the overlap between the predicted and ground truth segmentation masks as given in Eq. (2). It is especially important when dealing with imbalanced datasets where the glaucoma-affected regions may be small in comparison to the overall image. This balanced metric takes into account both precision (the ability to correctly classify positive instances) and recall (the ability to capture all positive instances).
The F1 Score ranges from 0 to 1, with higher values indicating better model performance in terms of both precision and recall.
iii. Sensitivity (True-Positive Rate or Recall): evaluates the extent to which the model identifies positive instances (glaucomatous regions) out of all positive instances as given in Eq. (3). It is the ratio of real positives to true positives and false negatives [37]. It pertains to the model's glaucoma detection. High sensitivity means the model can accurately identify glaucomatous regions, reducing false negatives. In medical applications like glaucoma diagnosis, false negatives can be serious, therefore sensitivity is important.
The Novel UNet-Inception Attention architecture uses these measures to evaluate the model's ability to segment glaucoma-affected retinal areas. These parameters measure the model's accuracy, Dice Score, and sensitivity in minimizing false positives and negatives [38], ensuring the glaucoma segmentation system's dependability and clinical relevance.
4.3 Results Evaluation
This study presents and rigorously validates a novel heuristic-based framework for segmentation and classification of glaucoma optic nerves. Utilizing the assets of both the UNet and Inception frameworks, this novel method automatically segments glaucoma-affected regions by integrating them in a seamless manner. In addition, it employs the Harris Hawks method for feature selection, which improves its accuracy, and a hybrid loss function to maximize performance.
The perception, which introduces non-linearity further into system, is the last step. Rectified linear units (ReLU) [39] is the most commonly employed perception in ML. To get the greatest result in retinal classification, the values used for each set of parameters were established through study and testing. Using a blended UNet-Inception structure with Opportunistic hawks optimization for selecting features, the suggested theory's loss and efficiency plot is displayed for training and validation sets for 30 iterations. The figure shows the trade-off that occurs between both the model's loss and correctness during learning, with marginally lower levels and highly accurate values indicating improved results.
The outcomes show that the suggested approach beats cutting-edge algorithms in accuracy and loss [40]. During the course of the 30 epochs, the validation error and accuracy for the suggested technique slowly get less and better, respectively, showing that the system was learning and getting better at what it does. In comparison to the most recent models, the suggested technique outperforms them with validating accuracy values of 0.932 and 0.9678 for ONH and RNFL segmentation, correspondingly.
Table 3: Class-wise Metrics after 5 fold cross-validation
Architecture
|
ONH Accuracy
|
RNFL Accuracy
|
ONH Sensitivity
|
RNFL Sensitivity
|
ONH Dice Score
|
RNFL Dice Score
|
DeepVessel-Net [19]
|
0.905 ± 0.16
|
0.876±0.13
|
0.879±0.19
|
0.844±0.21
|
0.872±0.09
|
0.827+0.14
|
DeepOptic-Net [17]
|
0.912 ± 0.11
|
0.882±0.14
|
0.887±0.17
|
0.856±0.24
|
0.887±0.09
|
0.837±0.14
|
Multi-Branch Network [21]
|
0.922 ± 0.07
|
0.893±0.14
|
0.898±0.19
|
0.864±0.27
|
0.902±0.07
|
0.847±0.08
|
U-Net Inception with Attention (proposed)
|
0.932 ±0.04
|
0.9678±0.02
|
0.9478±0.09
|
0.9554±0.12
|
0.918±0.07
|
0.9224±0.06
|
Table 3 summarizes the ONH and RNFL segmentation accuracy, sensitivities, and Dice scoring for each approach. In comparison to cutting-edge approaches, the suggested method delivers superior ONH and RNFL segmentation accuracy, sensitivities, and Dice scoring. This shows that the suggested strategy, which uses a composite UNet-Inception structure and Harrison hawks optimization for feature selection, is more precise and successful at segmenting glaucoma, enabling improved disease detection and surveillance. Furthermore, the validation error for the suggested approach is markedly lower than that for state-of-the-art systems, pointing to enhanced network confluence and durability throughout training. The overall comparison of proposed model is presented in Table 4.
Table 4: State of the art Comparison for Overall Segmentation
Architecture
|
Accuracy
|
Dice Score
|
Sensitivity
|
DeepOptic-Net [17]
|
0.92
|
0.84
|
0.78
|
DeepVessel-Net [19]
|
0.94
|
0.87
|
0.81
|
Multi-Branch Network [21]
|
0.9
|
0.8
|
0.75
|
U-Net
|
0.88
|
0.78
|
0.71
|
U-Net Inception with Attention (proposed)
|
0.95
|
0.96
|
0.93
|
The evaluation of the implementation [41] of the proposed system and the most recent algorithms for ONH and RNFL categorization are shown in comparison in Fig. 6. Each method's correctness, sensitivities, and Dice scoring are displayed on the plot so that their results may be directly compared. In terms of ONH and RNFL segmentation, the suggested model performs better than cutting-edge methods thanks to improvements in correctness, sensitivities, and Dice scoring. This shows that the suggested technique for segmenting glaucoma is more accurate and efficient when it uses a combination UNet-Inception structure with Harrison hawks feature selection optimization.
Fig. 7 contrasts the suggested model's overall segmentation findings with the actual segmentation for both ONH and RNFL. The selected features are visually compared in the image as shown in Fig. 8, enabling for a subjective assessment of the outcome. The suggested technique closely matches the regression coefficients segmentation regarding ONH and RNFL segmentation accuracy. This shows how the proposed approach efficiently produces precise and dependable segmentation accuracy for ophthalmology evaluation and tracking.
The anticipated and actual segmented masks are side-by-side compared in the graphic, enabling a qualitative evaluation of the recognition rate. The ocular comparison demonstrates the suggested method's precision and dependability in producing exact division.
4.4 Discussions
The integration of Haris Hawk Optimization (HHO) technique and the UNet-Inception Attention architecture (UNet-IAA) for ophthalmology image segmentation has shown several significant outputs:
- Improved Accuracy and Effectiveness: By combining the UNet and Inception frameworks, you leverage the strengths of both architectures. UNet is well-known for its excellent performance in image segmentation tasks, while Inception enhances feature extraction and captures complex patterns. This integration allows for more accurate and effective ophthalmic categorization.
- Enhanced Feature Selection: The HHO technique, integrated with UNet-Inception, provides a powerful feature selection mechanism. HHO's optimization process helps identify the most informative features or parameters within the network, optimizing the model's performance. This can lead to better feature representations for ophthalmology image analysis.
- Automated Optimization: HHO automates the process of optimizing the UNet-Inception architecture. Instead of manually fine-tuning hyperparameters, HHO iteratively refines the architecture's parameters to minimize the objective function. This automation saves time and ensures that the model is effectively fine-tuned for the specific ophthalmology segmentation task.
- Objective Fitness Evaluation: The intersection over union (IoU) metric, used as the fitness function in HHO, quantifies the overlap between the predicted mask and the ground truth mask in the test dataset. This metric provides a clear and objective measure of how well the segmentation model is performing, allowing the HHO algorithm to guide the model towards better solutions.
- Efficient Optimization: HHO optimizes the UNet-Inception architecture efficiently by exploring various combinations of model parameters and configurations. It balances the exploration of new solutions with the exploitation of promising ones, which can lead to faster convergence and better final solutions.
- Potential for Generalization: The integration of HHO with UNet-Inception can lead to the development of a model that generalizes well to a wide range of ophthalmology image segmentation tasks. The automated optimization process helps create a versatile model that performs effectively on different datasets and categories within ophthalmology.