Towards efficient glaucoma screening with modular convolution-involution cascade architecture

Residual multi-scale feature fusion module

The RMSFFM enhances feature extraction for glaucoma detection by integrating multi-scale information, accommodating varied pathological signs. Key components include:

• Multi-size convolutions: Applies 1 × 1, 3 × 3, and 5 × 5 kernels to input feature maps. 1 × 1 adjusts channel relationships, 3 × 3 balances local spatial data, and 5 × 5 captures broader spatial context.

• Feature concatenation: Combines convolution outputs along channels, merging multi-scale spatial details into a unified feature map for subsequent processing.

• Feature fusion: Uses a 1 × 1 convolution post-concatenation to integrate multi-scale data, reduce dimensionality, and prepare features for later network stages.

• Residual connection: Merges the original input with the fused output via a skip connection, preserving input data integrity, mitigating vanishing gradients, and supporting deep network training.

Residual hybrid convolutional-involutional module

The RHCIM is a composite framework that combines depth-wise and dilated convolutions with involution operations, supported by residual connections. This structure leverages the efficiency of depth-wise convolutions, the broadened receptive field of dilated convolutions, and the adaptive capabilities of involution. The integration of these hybrid convolutional and involutional techniques within RHCIM represents a key contribution of this work.

Dilated convolution sub-module

Figure 3 illustrates a typical convolution process. Imagine that an input’s feature maps have heights and widths represented by $H$ and $W$. These maps are symbolized as $X$ in a space defined by the product of height, width, and the number of input channels $C_i$. We refer to a set of $C_o$ convolution filters by $F\in {\mathbb{R}}^{C_o\times C_i\times K\times K}$, where each filter is a collection of $C_i$ kernels, each of size $K$ by $K$. Each kernel within a filter is represented by $F_k\in {\mathbb{R}}^{C_i\times K\times K}$, where $k$ ranges from 1 to $C_o$, and each kernel element is denoted by $F_{k,c}\in {\mathbb{R}}^{K\times K}$, with $c$ ranging from 1 to $C_i$. To derive the output feature maps, designated as $Y\in {\mathbb{R}}^{H\times W\times C_o}$, these convolution filters are systematically applied to the input maps. This process involves sliding the filters over the input and conducting element-wise multiplications followed by a summation, as defined in Eq. (1).

(1) $$Y_{i,j,k}=\sum _{c=1}^{C_i}\sum _{\left(u,v\right)\in \delta }{F}_{k,c,u+\lfloor K/2\rfloor ,v+\lfloor K/2\rfloor }{X}_{i+u,j+v,c}.$$

In this context, $\delta \in {\mathbb{Z}}^2$represents the set of offsets in the neighborhood considered during convolution around the position $\left(u,v\right)$, expressed as:

(2) $$\delta =\left[-\lfloor K/2\rfloor ,\dots ,\lfloor K/2\rfloor \right]\times \left[-\lfloor K/2\rfloor ,\dots ,\lfloor K/2\rfloor \right].$$

Within this convolution module, the dilation operation is executed by inserting zeros between filter elements. This dilation technique enables the network to encompass more pertinent information by expanding the receptive field of the filters. Consider a filter $F$ of size $K\times K$ applied to an input with a dilation rate $d$. Instead of covering a $K\times K$ area of the input, the filter now spans an area of $\left[K+\left(K-1\right)\left(d-1\right)]\times [K+\left(K-1\right)\left(d-1\right)\right]$, with $\left(d-1\right)$ zeros inserted between each original element in the filter. This does not change the number of parameters, as the inserted values are zeros, but it does allow the filter to aggregate information over a broader area of the input. Our experiment utilized a dilated convolution within the RHCIM, employing a filter size of 3 × 3.

Involution sub-module

There’s a known issue with convolution operations in neural networks, where channels can be overly dependent on each other, causing a lack of adaptability. To improve this, a new type of kernel, called an involution kernel $H\in {\mathbb{R}}^{H\times W\times K\times K\times G}$, has been designed to change the way transformations are handled in both the spatial and channel domains (Fig. 4). For each pixel $X_{i,j}\in R^C$ located at ( $i$, $j$), the involution kernel $H_{i,j,.,.,g}\in {\mathbb{R}}^{K\times K}\left(g=1,2,\dots ,G\right)$ is customized, although it remains consistent across channels. A feature map is then created from the input by applying these involution kernels through a series of multiplications and additions. The variable $G$ represents the number of groups, with each group utilizing the same involution kernel. A feature map (expressed in Eq. (3)) is then created from the input by applying these involution kernels through a series of multiplications and additions.

(3) $$Y_{i,j,k}=\sum _{\left(u,v\right)\in \delta }{H}_{i,j,u+\lfloor K/2\rfloor ,v+\lfloor k/2\rfloor ,\lceil kG/C\rceil }{X}_{i+u,j+v,k}.$$

In contrast to convolution kernels, the configuration of involution kernels $H$ is contingent on the input feature map $X$. Alignment of the output kernels with the input is achieved by generating involution kernels based on the original input tensor. Consequently, the kernel generation function $\varphi$ and the mapping function at every position ( $i$, $j$) can be formulated as follows:

(4) $$H_{i,j}=\varphi \left(X_{{\psi }_{i,j}}\right).$$

Here, ${\psi }_{i,j}$ represents the set of pixels that $H_{i,j}$ depends on. The kernel generation function $\varphi :{\mathbb{R}}^C\to {\mathbb{R}}^{K\times K\times G}$ with ${\psi }_{i,j}=\left\{\left(i,j\right)\right\}$ is outlined as presented in Eq. (5).

(5) $$H_{i,j}=\varphi \left(X_{i,j}\right)=W_1\sigma \left(W_0{X}_{i,j}\right).$$

In this context, $W_0$ belonging to ${\mathbb{R}}^{\frac{C}{r}\times C}$and $W_1$ belonging to ${\mathbb{R}}^{\left(K\times K\times G\right)\frac{C}{r}}$ represent two linear transformations that jointly form a bottleneck structure. The reduction in channels, governed by a ratio $r$, is utilized to enhance computational efficiency. Additionally, the symbol $\sigma$ signifies the inclusion of batch normalization and non-linear activation functions between the two linear projections. The process of feature generation described in Eq. (3) can be seen as an extended version of self-attention (Vaswani et al., 2017). In self-attention, values $V$ are pooled according to affinities determined by calculating the similarity between the query $Q$ and the key $K$, as expressed in Eq. (6).

(6) $$Y_{i,j,k}=\sum _{\left(p,q\right)\in \delta }{\left(Q{K}^T\right)}_{i,j,p,q,\lceil kH/C\rceil }{V}_{p,q,k}.$$

In this scenario, $Q$, $K$, and $V$ experience linear transformations from the input $X$, while $H$ represents the number of heads in multi-head self-attention (Vaswani et al., 2017). The similarity arises from both operators gathering pixels within the neighborhood $\vartheta$ or a more loosely bounded range via a weighted sum. Regarding involution computation, it can be interpreted as a form of spatial attentive aggregation. Conversely, the attention map (also termed affinity or similarity matrix) $Q{K}^T$ in the self-attention mechanism can be seen as a variant of the involution kernel $H$.

In Fig. 4, the involution kernel $H_{i,j}\in {\mathbb{R}}^{K\times K\times 1}$ (where $G=1$ for simplicity in demonstration) is generated from the function $\varphi$ utilizing a single pixel at position ( $i$, $j$), after which a rearrangement from channel to space is performed. The involution’s multiply-add operation encompasses two stages: $N$, indicating multiplication broadcasted across $C$ channels, and $L$, signifying summation aggregated within the $K\times K$ spatial neighborhood.

Advanced feature analysis with LightGBM

In the proposed MCICNet framework, the extraction of high-level feature maps is a crucial step for the subsequent classification task. After passing through the various layers of the MCICNet architecture, the network generates feature maps that encode essential spatial and contextual information from the input image. These feature maps are then flattened into a one-dimensional vector for each input sample, preparing them for the classification stage. Unlike conventional CNNs that rely on FC layers for classification, the proposed MCICNet employs LightGBM as an advanced classifier to handle the extracted feature vectors. This choice is motivated by LightGBM’s ability to efficiently manage high-dimensional data while maintaining accuracy (Ke et al., 2017).

LightGBM builds decision trees sequentially, with each tree correcting errors from the previous ones. It initializes predictions to zero and iteratively refines them through boosting rounds. In each round, residuals (differences between true labels and current predictions) are calculated, and a new decision tree is trained to predict these residuals. Predictions are updated by adding a fraction of the tree’s output, scaled by a learning rate. After multiple rounds, final predictions are determined using a threshold function. This approach allows MCICNet to handle complex classification tasks efficiently, avoiding the computational overhead of FC layers. The hybrid MCICNet-LightGBM model is outlined in Algorithm 1, available in the Supplemental Files.

Comparative design framework

To provide a comprehensive evaluation of the MCICNet model’s performance, this section details the methodology employed to systematically compare it with established models.

For comparative analysis, nine prominent CNN architectures were chosen as benchmarks against MCICNet: SqueezeNet, AlexNet, MobileNetV2, DenseNet121, ResNet18, GoogLeNet, ShuffleNet, EfficientNetB0, and VGG16. Additionally, three transformer-based architectures were evaluated for their performance in this glaucoma detection task: Vision Transformer (ViT), MaxViT, and Swin Transformer (SwinT).

For consistent evaluation, all models were tested with an input size of 224 × 224, except for SqueezeNet, which required an input size of 227 × 227. This input size is commonly used in image classification tasks and ensures compatibility across the models. All models were trained using the Adam optimizer with a learning rate of 0.001. No hyperparameter optimization was conducted in this study, and the models were trained with their default architectural settings, aside from the specified optimizer and learning rate, to maintain consistency in the comparison.

To test the hybrid Involution-Convolution method in MCICNet, a structurally identical model replaced involution layers with convolution layers in the RHCIM, creating MCICNet-NoInvolution. Filter numbers and kernel sizes aligned with the dilated convolution setup. This examined the benefits of involution’s adaptive, spatially sensitive kernels versus convolution’s fixed kernels in capturing complex spatial data patterns.

Dataset description and preprocessing techniques

This study used the Large-scale Attention-based Glaucoma (LAG) dataset, containing 5,824 retinal fundus images (2,392 glaucoma, 3,432 healthy) from Beijing Tongren Hospital. Standardized at 500 × 500 pixels, the JPEG images were evaluated by glaucoma experts using morphological (OD, RNFL) and functional (IOP, visual field) assessments, with manual OD reviews for diagnostic accuracy. Each image was labeled as glaucoma or non-glaucoma. A subset of 4,854 images is available upon request, and the dataset can be downloaded at https://github.com/smilell/AG-CNN.

To further assess the effectiveness of the proposed model, an additional dataset, ACRIMA, was utilized (Diaz-Pinto et al., 2019). This dataset consists of 705 retinal fundus images, comprising 396 glaucomatous cases and 309 normal healthy images. These images were acquired from both the left and right eyes and were previously dilated and centered on the OD. The retinal images have a resolution of 2048 × 1536 pixels. The dataset was examined and annotated by two ophthalmologists at the Foundation Ophthalmological Mediterranean. The ACRIMA dataset is available for download at https://figshare.com/s/c2d31f850af14c5b5232.

Both datasets were partitioned into three subsets to ensure an effective training and evaluation process. Specifically, 70% of the data was allocated to the training set for optimizing the model’s parameters, while 15% was designated as the validation set to monitor performance and guide model adjustments during training. The remaining 15% was reserved as the test set to evaluate the model’s generalization on previously unseen data. It is important to note that the 70:15:15 split is at the image level and is a standard practice commonly used in ML. This split ratio is chosen to ensure a balanced distribution of data for training, validation, and testing, which is crucial for training robust models and evaluating their performance accurately. The 70% training set provides a sufficiently large dataset to train the model effectively, while the 15% validation and test sets offer adequate samples to monitor and evaluate the model’s performance, ensuring that the model generalizes well to unseen data.

Moreover, to enhance model training stability and consistency, pixel normalization was performed on each image, scaling pixel values from the original range of 0 to 255 to a range of 0 to 1 by dividing each pixel by 255.0. This normalization step aids the model in processing image data more efficiently by mitigating variations in brightness and contrast, thus promoting optimal network performance and improving generalization across the dataset. Additionally, all images were resized to 96 × 96 pixels to reduce computational demands, while maintaining sufficient resolution for model training. Both MCICNet and MCICNet-NoInvolution were tested with this smaller input size, establishing a model structure that is computationally efficient without compromising performance. Lastly, image enhancement and data augmentation techniques were not applied to preserve the original characteristics of the images, ensuring a more realistic assessment of the model’s performance on raw data.

Evaluation metrics

The proposed model’s performance is assessed quantitatively using metrics such as accuracy, sensitivity, precision, specificity, and F1-score, as outlined in Eqs. (7)–(11):

(7) $$accuracy={\displaystyle \frac{TP+TN}{FP+FN+TP+TN}}$$ (8) $$sensitivity={\displaystyle \frac{TP}{TP+FN}}$$ (9) $$precision={\displaystyle \frac{TP}{TP+FP}}$$ (10) $$specificity={\displaystyle \frac{TN}{TN+FP}}$$ (11) $$F1=2\times {\displaystyle \frac{PREC\times SEN}{PREC+SEN}.}$$

Here, true positives (TP) correctly identifies glaucoma cases, false positives (FP) misclassifies healthy individuals as glaucomatous, false negatives (FN) misses glaucoma cases, and true negatives (TN) accurately confirms non-glaucoma cases. Additionally, the Matthews correlation coefficient (MCC) is included as a more balanced metric, particularly useful when dealing with imbalanced datasets. The MCC is defined as:

(12) $$MCC={\displaystyle \frac{TP\times TN-FP\times FN}{\sqrt{\left(TP+FP\right)\left(TP+FN\right)\left(TN+FP\right)\left(TN+FN\right)}}.}$$

Experimental environment and parameter setting

The training setup for glaucoma detection was established on a dual NVIDIA Tesla T4 GPU configuration, each offering 15 GB of memory, supplemented by 29 GB of system RAM to maintain stable and efficient training processes. All models were implemented and trained using Python version 3.10. TensorFlow with Keras was employed for constructing the feature extraction component of MCICNet, while the LightGBM algorithm was implemented using the Scikit-learn library.

LightGBM parameters were tuned using GBDT for pattern capture. A 0.1 learning rate, 200 boosting rounds, and 31 leaves per tree balanced performance and generalization. Tree depth was unrestricted, with 20 samples minimum per leaf to curb overfitting. Feature and bagging fractions were 1.0 for full data use, and class weights auto-adjusted for imbalance.

For the additional benchmark models, pre-trained weights from ImageNet were utilized to support TL. Input images were resized according to each model’s specific input dimensions and normalized based on ImageNet’s mean and standard deviation. The Adam optimizer was applied for weight updates during training, with a batch size of 32 over a training span of 200 iterations. Additionally, a fixed random seed (set to 7) was used to ensure the reproducibility of the training results.

Convergence analysis

Figure 5 illustrates the progression of training and validation accuracy throughout the training process for MCICNet, MCICNet-NoInvolution, AlexNet, MobileNetV2, SqueezeNet, ResNet18, GoogLeNet, DenseNet121, EfficientNetB0, ShuffleNet, VGG16, ViT, MaxViT, and SwinT. A significant observation at the 200th iteration is the superior performance of MCICNet, achieving a final accuracy of 96.2%, followed closely by MCICNet-NoInvolution at 95.6%. This highlights the substantial improvement in learning efficacy achieved through the incorporation of involution operations within the network. In contrast, all other models, including high-performing architectures such as VGG16, recorded final accuracies between 86.4% and 95%. Among the transformer-based models, ViT, MaxViT, and SwinT demonstrated relatively moderate performance, with final validation accuracies below 93%. ViT outperformed MaxViT and SwinT, likely due to its more effective handling of spatial dependencies, while MaxViT’s reliance on feature pooling and SwinT’s hierarchical approach may not have aligned as well with the dataset characteristics. ViT also surpassed most CNN-based models, such as SqueezeNet, ResNet18, and GoogLeNet, except for VGG16 and the two variants of MCICNet. This performance difference can be attributed to ViT’s attention mechanism, which allows the model to focus on relevant regions of an image, thereby improving performance, particularly when global context is important.

With regard to stability and convergence, MCICNet demonstrates smoother validation accuracy curves with minimal fluctuations compared to models such as MobileNetV2 and GoogLeNet, which exhibit greater variability across epochs. This stability suggests that MCICNet is resilient to training fluctuations, enhancing its likelihood of rapid and reliable convergence. Furthermore, MCICNet’s fast convergence over epochs suggests an efficient learning process, potentially due to its multi-scale feature fusion and hybrid convolution-involution architecture, which effectively captures pertinent features in fundus images. Conversely, models such as SqueezeNet and ShuffleNet demonstrate a slower learning progression, indicating that they may require additional epochs to achieve optimal performance or may not be as well-suited to the glaucoma detection task as MCICNet.

Several models, including AlexNet, MobileNet, ResNet, GoogLeNet, DenseNet, EfficientNet, ShuffleNet, VGG16, ViT, MaxViT, and SwinT, exhibit a noticeable disparity between training and validation accuracies, indicating overfitting. In these cases, the models perform substantially better on the training data than on the validation data. In contrast, models such as MCICNet-NoInvolution and MCICNet-Involution demonstrate minimal overfitting, as their training and validation accuracies are closely aligned, suggesting improved generalization to unseen data. Certain models, particularly those more prone to overfitting, display fluctuations in validation accuracy, which indicates heightened sensitivity to the specific characteristics of the training data and a reduced ability to generalize to new data.

Performance evaluation

To illustrate the effectiveness of the proposed MCICNet, the metrics precision, sensitivity, specificity, F1-score, accuracy, MCC, and AUC were calculated for both the validation and test datasets, as presented in Tables 1 and 2, respectively. The proposed MCICNet demonstrates superior performance, achieving the highest accuracy (96.29% on the validation set and 95.47% on the test set) and AUC (0.9933 on the validation set and 0.9914 on the test set) among all models, including both older and newer architectures. This is especially significant given that MCICNet is a novel architecture designed specifically for glaucoma detection, utilizing a combination of convolution and involution operations. In comparison, MCICNet-NoInvolution, which replaces involution layers with traditional convolution layers, also performs well but slightly underperforms relative to MCICNet. This indicates that the inclusion of involution operations enhances the model’s ability to capture cross-channel dependencies and spatial features, thus improving performance. Among traditional CNN models, VGG16 ranks as the second-best performer, despite being the oldest model in the comparison after AlexNet. VGG16 achieves 95.06% accuracy on the validation set and 93.55% on the test set, outperforming more recent architectures such as ResNet, DenseNet, SqueezeNet, and EfficientNet. This result is noteworthy, as it is somewhat unexpected for older models like VGG16 to surpass newer architectures. The high accuracy of VGG16 suggests that certain architectural elements may still be relevant in glaucoma detection, even when compared to more recent innovations in CNN design. Transformer-based models, including ViT, MaxViT, and SwinT, demonstrate moderate performance, with SwinT achieving the highest test accuracy (93.55%) among the three. Despite their success in natural language processing and certain computer vision tasks, transformers’ performance in glaucoma detection does not match that of MCICNet. One possible explanation for this is that transformer models typically require large-scale datasets to realize their full potential. While the LAG dataset is substantial, it may not be sufficient to exploit the complete capabilities of transformers, limiting their performance for this specific task. The generalization ability of a model, assessed through its performance consistency across validation and test sets, is crucial for real-world applicability. MCICNet exhibits strong generalization, with only a slight performance drop between the validation and test datasets. Specifically, the accuracy decreases from 96.29% to 95.47%, and the AUC drops from 0.9933 to 0.9914. These minimal reductions suggest that MCICNet is robust and can maintain high accuracy on unseen data, a valuable attribute for clinical applications where reliable and consistent results are essential.

ROC and AUC analysis

ROC curves were employed to assess the efficacy of the proposed MCICNet on both the validation (Fig. 6A) and test sets (Fig. 6B). The ROC curve graphically represents the relationship between TPR and FPR across various threshold values. On the validation set, MCICNet achieved a high AUC of 0.9933, with its ROC curve closely following the top-left boundary of the ROC space, signifying a robust TPR with minimal FPR at multiple threshold points. Similarly, MCICNet-NoInvolution also demonstrated strong performance with an AUC of 0.9907, only marginally lower than MCICNet with involution. Although the VGG16 architecture, with its substantial parameter count, achieved a competitive AUC of 0.9879, it was outperformed by the MCICNet models. Noteworthy models also include EfficientNetB0 and ShuffleNet, both of which yielded AUC values above 0.97, whereas AlexNet exhibited lower efficacy (AUC = 0.9547). On the test set, the proposed MCICNet continued to demonstrate leading performance, achieving an AUC of 0.9914, underscoring its generalizability to previously unseen glaucoma images. Similarly, MCICNet-NoInvolution maintained a high AUC of 0.9887, highlighting the effectiveness of the modular cascading design. While the remaining models displayed stable performance, they did not match the MCICNet architectures in accuracy.

External dataset evaluation with seed averaging for stability

To validate the robustness and generalizability of the MCICNet model, an external evaluation was conducted using the ACRIMA dataset. This step was critical to confirm that MCICNet’s performance was not solely dependent on the characteristics of the LAG dataset used for training and initial validation. Figure 7 illustrates the accuracy and loss curves for the training and validation phases on the ACRIMA dataset. Here, MCICNet showed commendable performance; the training and validation accuracy curves in Fig. 7A quickly converged to high levels, with validation accuracy closely tracking the training curve, indicating no significant overfitting. The training accuracy neared 100%, while the validation accuracy stabilized at 97.14%, demonstrating good generalization to new data. The loss curves in Fig. 7B similarly declined sharply and then stabilized, with minimal divergence between training and validation losses, further attesting to the model’s robustness.

The stability and generalizability were further explored through seed averaging experiments on the LAG and ACRIMA datasets, with findings reported in Tables S1 and S2. These experiments used two different seeds (Seed A and Seed B) for model initialization, aiming to ensure performance consistency despite random weight initialization. For the LAG dataset, MCICNet achieved an average accuracy of 95.61% with negligible variance between seeds. On the ACRIMA dataset, accuracy was consistent across seeds as well. This stability across different metrics like precision, sensitivity, specificity, and AUC underscores MCICNet’s reliability for medical imaging applications.

A notable aspect of this study is the model’s capability to generalize across datasets with different characteristics. The ACRIMA dataset, unlike LAG, features pre-cropped OD regions instead of full fundus images. Despite this, MCICNet delivered an average accuracy of 96.26%, precision of 99.34%, and sensitivity of 96.09%, indicating strong adaptability.

MCICNet consistently outshone other models across both datasets. On the LAG dataset, it achieved a superior specificity of 95.11% compared to VGG16 and transformer models like ViT. On ACRIMA, it achieved perfect specificity at 97.50%, contrasting with lower sensitivity in other models. This performance can be attributed to MCICNet’s hybrid architecture, combining convolution with involution to capture both local and global features, and the use of LightGBM for classification, which enhances decision-making in distinguishing glaucomatous cases. Conventional CNNs and transformers, while effective, have limitations in handling smaller datasets or capturing long-range dependencies, areas where MCICNet proves advantageous. Its success on relatively smaller datasets reaffirms its suitability for glaucoma detection in clinical settings.

Parameter count, training, and inference time comparison

Computational efficiency was assessed by comparing parameter counts (Fig. 8). VGG16, with 138 million parameters, demands significant resources, while SqueezeNet (1.2 million) and ShuffleNet (2.2 million) are more efficient. MCICNet, with just 0.9 million parameters, surpasses MCICNet-NoInvolution and other pretrained models, owing to its involution-convolution integration, suiting resource-limited settings and enabling fast inference for glaucoma detection.

Training and inference times (Table 3) highlight MCICNet’s efficiency. On the LAG dataset, MCICNet trained in 2 min 2 s, far quicker than VGG16 (15 min 15 s) and ViT (16 min 14 s), reflecting faster convergence. On ACRIMA, it took 33 s. Inference times for LAG (729 images) ranged from 0.0037 s (SwinT) to 0.0210 s (MCICNet-NoInvolution), and for ACRIMA (107 images), from 0.005 ms (SwinT) to 0.029 ms (MCICNet-NoInvolution). MCICNet’s inference time of 0.0197 s, though higher due to LightGBM, remains clinically viable, balancing speed and accuracy effectively for medical imaging tasks.

Performance robustness under noise and low contrast

This section examines the robustness of the MCICNet model by testing its performance on retinal fundus images from the LAG and ACRIMA datasets that have been altered with Gaussian noise and low contrast. These modifications mimic real-world imaging challenges that could compromise diagnostic accuracy. Gaussian noise introduces random pixel intensity fluctuations, simulating poor image quality, whereas low contrast reduces the range of pixel intensities, making images appear less distinct. The performance of MCICNet under these conditions is detailed in Table 4. On the LAG dataset, accuracy slightly dropped from 95.47% in original conditions to 94.78% with Gaussian noise and 95.19% in low contrast scenarios. The AUC also experienced a small decline, going from 0.9914 to 0.9870 with noise and to 0.9890 in low contrast. Despite these decreases, MCICNet maintained a high level of performance, demonstrating its robustness. For the ACRIMA dataset, the model showed even stronger resilience. With Gaussian noise, accuracy reduced marginally from 96.26% to 95.32%, but the AUC stayed robust at 0.9887. Interestingly, under low contrast conditions, MCICNet’s performance actually improved, with accuracy increasing to 97.19% and AUC to 0.9933. This suggests that MCICNet is particularly adept at managing low-contrast images, which is beneficial in clinical settings where lighting and camera settings can vary, leading to inconsistent image quality.

Classifier comparison for MCICNet

The performance of LightGBM in MCICNet was evaluated against various classifiers—RF, SVM, KNN, logistic regression (LR), extra trees (ET), NB, decision tree (DT), stochastic gradient descent (SGD), AdaBoost, XGBoost, and an FC-layer MCICNet variant—using the LAG and ACRIMA datasets (Table 5). On LAG, LightGBM achieved top results: 95.47% accuracy, 95.15% precision, 94.91% sensitivity, 95.03% F1-score, and 0.9914 AUC, outperforming others due to its gradient-boosting strength in handling complex features. On ACRIMA, AdaBoost edged out in accuracy (97.19% vs. 96.26%), but LightGBM led in precision (96.47%), sensitivity (95.98%), F1-score (96.19%), and AUC (0.9943), showing adaptability across datasets. XGBoost trailed closely on LAG, aided by robust tree-building, while SVM and LR lagged in precision and F1 due to linear limitations. Ensemble methods (RF, ET) performed well but fell short of LightGBM’s precision. NB and DT struggled with complexity and overfitting. Compared to FC layers, LightGBM’s iterative focus on difficult samples enhanced data insight over linear aggregation.

Model predictions and feature map analysis

To explore MCICNet’s performance in glaucoma detection, its predictions and feature extraction were analyzed using retinal fundus images. Examples of model predictions were visualized in Fig. 9 to assess classification, showing MCICNet’s high accuracy in distinguishing glaucomatous from non-glaucomatous cases—correct predictions in green, misclassifications in red. It excels even with poor OD visibility, outperforming traditional CNNs and advanced methods, with rare errors linked to degraded retinal structures. Further insight into MCICNet’s internal workings was gained by examining feature maps at different network layers, as shown in Fig. 10. This visualization reveals how the model processes images, with high activation areas (in red) indicating regions crucial for prediction, while low activation areas (in blue) show less influence on the outcome. The analysis shows that activation is concentrated around the OD, underscoring the model’s focus on critical areas for glaucoma detection, like the RNFL. The model’s use of both convolution and involution operations allows it to capture both local and global features, enhancing its ability to detect subtle changes indicative of glaucoma. The involution operation’s adaptability in generating kernels based on input maps directs attention to the most informative parts of the image, aiding in pinpointing disease progression. The findings from this feature map analysis, coupled with the model’s predictions, validate MCICNet’s effectiveness in glaucoma detection. It demonstrates that MCICNet not only captures relevant features with precision but also maintains high performance under challenging conditions. The strategic concentration of activation on clinically significant regions reinforces the model’s reliability.