A Pyramid-Based Feature Fusion Framework for Keratoconus Detection Using LightGBM

Keratoconus is a progressive, bilateral disorder [1] of the cornea characterized by thinning and bulging of the corneal tissue, resulting in conical shape of the cornea [2]. This change in the corneal shape leads to irregular astigmatism and vision impairment [3]. The disease usually manifests during adolescence and progresses until the fourth decade of life [4]. Unlike the early stages of the disease, moderate to advanced stages are easily diagnosed due to the presence of classic signs observable via slit-lamp and clinical examination [5]. The early detection of the disease is crucial as progression can be halted using medical intervention [6]. However, the detection of the disease in its incipient stage is often challenging as the early signs are often subtle and not easily detectable through conventional methods [7].

In recent years, Machine Learning (ML) approaches have shown promise as an automated diagnostic tool in healthcare [8]. These techniques offer objective, data-driven classification by identifying subtle patterns and abnormalities, potentially enhancing diagnostic accuracy and aiding healthcare professionals when integrated into automated screening systems [9].

In this paper, multi-scale representation of corneal maps using image pyramids is proposed. Histogram of Oriented Gradients (HOG) is then applied to extract features across these scales of corneal maps in L*a*b* colour space. Gradient based features like HOG measure the rate and direction of change in an image [10]. Since, Keratoconus is defined by a change in corneal shape, HOG descriptors are well suited to detect these changes in cornea. At the fine scale, HOG captures localized structural changes such as steepening and surface irregularities, which are often indicative of early-stage Keratoconus. At the coarse level, it highlights broader deformation patterns, including generalized thinning and changes in corneal curvature that are characteristics of the advanced disease [11]. L1-penalized Logistic Regression is used for feature selection and hyperparameter tuned LGB classifier is used to evaluate the classification performance of the model.

The main aim of this study is to develop an automated multiscale image analysis framework for keratoconus detection from corneal topographic maps using multiscale image representation and feature extraction via image pyramids and Histogram of Oriented Gradient (HOG) Descriptors and LightGBM for classification.

The structure of this paper is outlined as follows. Section 2 reviews previous studies on the detection of Keratoconus, focusing on applications of machine learning techniques as well as image analysis approaches based on pyramid structures and HOG in medical imaging. Section 3 describes the proposed methodology in detail, covering aspects such as dataset collection and pre-processing, imbalance handling, multi scale representation of images using Gaussian, Laplacian and Patch-based pyramids, feature extraction using HOG Descriptors in L*a*b* color space, classification strategy using LightGBM classification framework and statistical analysis. Section 4 discusses experimental setup, performance evaluation methods and results. A thorough analysis and interpretation of the findings are provided in section 5. Finally, section 6 summarizes the main conclusions of this study and offers suggestions for future research directions.

3.1 Data collection and pre-processing

In this work, a total of 2961 corneal images obtained from 542 eyes using the Oculus Pentacam imaging system is utilized. This dataset, which had been previously compiled and made publicly accessible through earlier publications [17], includes three categories: Keratoconus, Normal and Suspect. Image labeling was performed by three experienced corneal specialists following established guidelines as reported in prior studies [17].

11.png

12.png

13.png

Specifically, the dataset comprises of 1050 images of each Keratoconus and Normal categories, while the Suspect group contains 861 images. For model development and evaluation, the entire dataset is partitioned into stratified split of 70% for training and 15% each for validation and testing. This resulted in 2072 images in the train set, 444 images in the validation set and 445 images in the test set. The size of each corneal map is 224×224×3. Figure 1. shows the corneal topographic maps of keratoconic, normal and suspect eyes derived from Pentacam device.

To address the issue of class imbalance within the dataset, class weights are computed using a balanced weighting strategy. This technique assigns proportionally higher weights to classes with fewer samples, thereby promoting fairer learning during the model training. The weight for each class $w_j$ is determined using the formula:

$w_j=\frac{N_{\text {train }}}{C \times N_j}$            (1)

where,

$w_j$ represents the weight assigned to class $j$,

$N_{\text {train }}$ is the total number of images in the train set,

$C$ is the total number of categories, and

$N_j$ is the number of training images belonging to class $j$.

Besides class weighting, Synthetic Minority Over-Sampling Technique (SMOTE) is also employed to handle data imbalances. SMOTE [24] is used to equalize the number of samples in the suspect class to that of normal and keratoconus classes in the training set. Validation and test sets are left untouched. An extra of 133 synthetic samples from suspect class are generated to balance the dataset.

To enhance the discriminative quality of the corneal maps, all images are converted from RGB to L*a*b* color space prior to feature extraction. This conversion is inspired by prior work [20].

3.2 Multi-scale representation using Image pyramids

To capture multi-scale features from corneal topography, three distinct pyramid-based image representations were constructed. Gaussian pyramids, Laplacian pyramids [25] and patch-based pyramid [22]. Each image was first converted to the L*a*b* colour space to enhance the corneal maps. The Gaussian pyramid is constructed by iteratively applying Gaussian smoothing followed by a downsampling with factor of 2 at each level. A total of 3 Gaussian levels is generated beyond the original image. A Laplacian pyramid is derived by subtracting the upsampled lower-level Gaussian image from its corresponding higher-level image, emphasizing localized edge and texture details at each size. Three Laplacian levels corresponding to the three Gaussian levels are computed. Additionally, a patch-based image pyramid is constructed by directly resizing image with downsampling factor of 2 per level, without applying any smoothing operation. At each resolution, the downscaled image is partitioned into non-overlapping 28×28 patches.

3.3 Histogram of Oriented Gradients

The Histogram of Oriented Gradients (HOG) is a feature descriptor in computer vision and image analysis tasks [26]. HOG features are independently extracted from each pyramid type to capture texture and edge information. For both the Gaussian and the Laplacian pyramid, HOG features are computed from each image level independently across all scales. In the case of the patch-based pyramid, each downscaled image is subdivided into non-overlapping patches of size 28×28 pixels, and HOG features are extracted individually from each patch. A stride of 28 pixels is used to ensure complete spatial coverage without overlapping patches.

Table 1 presents the HOG descriptor parameters used for feature extraction at each pyramid level.

The parameters in Table 1 in are chosen in a way to strike a balance between capturing discriminative features and maintaining computational efficiency. 7 orientation bins were chosen to capture major edge directions while keeping computational cost to a minimum. 4x4 pixels per cell is chosen to capture finer details when compared to the default of 8×8 pixels per cell. A 2×2 cells per block was found to be the most effective in capturing local variations and small enough to not average out the important detail. L2-norm with hysteresis is chosen as it prevents few high contrast gradients from dominating the entire feature vector [10].

Table 1. HOG descriptor parameters

3.4 Feature selection

To reduce the feature dimensionality and retain only the most discriminative features, a coordinate descent based L1- regularized Logistic Regression is employed for feature selection with solver set to ‘liblinear’, max iterations are set to 10000 and the parameter ‘C’ is set to 1.0. L1 penalty encourages sparsity by driving the coefficients of less informative features to zero effectively selecting a subset of relevant features for classification [21]. The model is trained on scaled HOG feature vectors extracted from each pyramid representation, and the selected features are subsequently used as input to the final LightGBM classifier. This approach improved computational efficiency and reduced the risk of overfitting in high dimensional feature spaces.

3.5 LightGBM classifier

LightGBM [27] is an efficient gradient boosting framework designed for machine learning applications. Unlike conventional gradient boosting models that use a level-wise approach, LightGBM adopts a leaf-wise tree growth strategy. In this approach, the leaf with the highest potential gain is selected and split at each iteration, enabling the model to focus on the most informative regions of the feature space. The selected features by the L1-Logistic Regression are used as input to the LightGBM classifier. To identify the optimal hyperparameters for the LightGBM classifier, a random search cross-validation (RandomizedSearchCV) strategy is employed. This approach samples a predefined number of random combinations from specified hyperparameter distributions, offering a more computationally efficient alternative. In this study, 30 randomly selected hyperparameter configurations are evaluated for each feature extraction pipeline.

Table 2 presents description of the hyperparameters employed in the model training process along with their respective search ranges utilized during optimization.

To maintain data integrity and prevent overlap between training and validation subsets, a PredefinedSplit (ps) is utilized during cross-validation. The model selects the best performing hyperparameter set after training on the train set based on the validation accuracy, evaluated on a reserved validation set containing 444 samples. Once the best configuration of the hyperparameter set is found, the model’s classification performance is evaluated finally on the test set containing 445 samples. The workflow for proposed machine learning framework for keratoconus classification is depicted in Figure 2.

Table 2. Hyperparameters and search ranges for model optimization

2.png

Figure 2. Workflow of the proposed machine learning framework for keratoconus detection

31.png

(a) Gaussian pyramid constructed using 3 downsampling steps resulting in 4 levels with resolutions: 224×224, 112×112, 56×56 and 28×28

32.png

(b) Laplacian pyramid constructed using 3 downsampling steps resulting in 3 levels with resolutions: 224×224, 112×112, and 56×56

33.png

(c) Patch-based pyramid constructed using 3 downsampling steps without smoothing resulting in 4 levels with resolutions: 224×224, 112×112, 56×56 and 28×28

Figure 3. The Gaussian, Laplacian and Patch-based pyramids

Furthermore, bootstrapping with n = 1000 resamples is employed to account for variability. For each resample, models are trained and evaluated. The results of each iteration are recorded and stored in .csv files. A one-way Analysis of Variance (ANOVA) is performed across the three pyramid-HOG pipelines to test whether the mean performance metrics differed significantly. Tukey’s Honestly Significant Difference (HSD) post-hoc test is used for pairwise comparison between the three pyramid-HOG pipelines that are statistically significant. Pairwise t-tests are conducted on bootstrap derived distributions to compare class weighting and SMOTE imbalance handling. These statistical methods are used in machine learning research to compare model performance [28]. Figure 3 shows the corneal maps represented at different scales using Gaussian, Laplacian and Patch-based pyramid.

In this study, an automated machine learning framework for classification of Keratoconus using image pyramids, HOG and LightGBM classifier is proposed. A stratified split of 70%-15%-15% of train validation and test split is employed on a dataset of 2961 pentacam derived corneal maps containing keratoconus, normal and suspect classes. The dataset is imbalanced with the “suspect” class having relatively fewer instances. To address this, class weighting based on inverse class frequency is applied during training, ensuring that minority class errors are penalized more heavily during optimization. RGB to L*a*b* conversion is used as the preprocessing step.

Three distinct multi-scale feature extraction approaches are investigated: Gaussian pyramid-HOG Laplacian Pyramid and a Patch-based Pyramid, with 3 levels each and downsampling factor of 2 per each level. Each technique aims to enhance feature representation by decomposing spatial information at multiple levels for detecting keratoconus. These decompositions are followed by HOG feature extraction, which captured the relevant discriminative features in the image. These extracted features are scaled using ‘StandardScaler’ and LightGBM classifier is trained with hyperparameter optimization via randomized search. From Table 3. It is evident that Gaussian Pyramid and HOG required a relatively higher learning rate and fewer leaves suggesting its multi-scale smoothed features resulted in a simple decision boundary. In contrast, the patch-based method required a ‘max_depth’ of 9 with a lower learning rate. All simulations are performed on Pop!_OS 22.04 (Ubuntu based Linux distribution) with 32GB of RAM and Ryzen 5600H Hexacore CPU. Model evaluation metrics in Table 4. revealed that all three pyramid strategies achieved strong classification performance, particularly for the keratoconus class, where precision and recall scores are consistently high. For instance, the patch-based pyramid model achieved a recall of 0.9304 for keratoconus class. However, suspect class detection remained challenging across all pipelines with precision ranging from 0.6863 to 0.7426. This underperformance is likely due to the intermediate morphological characteristics of suspect cases, which share traits with both normal and keratoconus corneas. Nevertheless, the patch-based pyramid exhibited slightly improved performance for suspect class classification, likely due to its ability to focus on localized regions where early ectatic changes are or likely to appear.

41.png

42.png

43.png

52.png

61.png

62.png

Table 5 provides the macro and weighted average metrics for each model. The Gaussian Pyramid and HOG pipeline delivered the best overall performance with a weighted average F1 score of 0.8581 and macro average AUC of 0.9568. Although, patch-based method trailed slightly in macro and weighted averages, it maintained a balanced performance across all classes, suggesting improving robustness in handling class imbalance. These results emphasize the effectiveness of image pyramids in enhancing class separability in corneal maps and validate the suitability of HOG descriptors in keratoconus classification. Figures 4, 5 and 6 show the confusion matrix plot and AUROC plot for all the three pipelines.

Figure 7 provides feature maps of corneal images across different scales and color channels. These feature maps confirm that model’s decisions are based on gradient-based information extracted by HOG.

71.png

Figure 7. The feature maps of the corneal image at multi scales

Table 6 provides statistical evidence for the performance differences observed in Tables 4 and 5. The F-statistic shown for each metric is very high (all are over 739). The high F-statistic shows that the probability of average performance of the three pyramid-HOG pipelines being the same is very low. The high F-statistic coupled with p-value <0.001 indicates that the differences are significant. Tukey’s HSD post-hoc test confirms that the difference between each pair of pipelines (Gaussian-HOG vs. Laplacian-HOG, Gaussian-HOG vs. Laplacian-HOG and Laplacian vs. Patch-based-HOG) is also statistically significant with adjusted p-values < 0.001. The positive mean difference indicates patch-based-HOG outperformed Laplacian-HOG in Laplacian vs. Patch-based HOG. The negative mean difference in Gaussian-HOG vs. Laplacian-HOG and Gaussian-HOG vs. Patch-based-HOG indicates that Gaussian-HOG has outperformed both the other methods.

Table 7 presents the results of a pairwise t-test comparing the performance of Class-Weighting and SMOTE within each of the three pyramid-HOG pipelines. This analysis statistically validates the effectiveness of each imbalance handling technique. The negative mean difference between Class Weighting and SMOTE indicates that Class Weighting has performed significantly better than SMOTE. The positive mean difference indicates that SMOTE has outperformed Class Weighting technique. For the Gaussian pyramid-HOG SMOTE slightly outperforms Class Weighting across most metrics with a positive mean difference with the exception of AUC-Macro. The p-values are less than 0.001, showing that the small differences are significant. In the Laplacian pyramid-HOG the difference between the two methods is very little. The p-value for accuracy metric is > 0.05 indicates that mean difference in accuracy is not significant. For all other metrics, the p-values are < 0.001. For the patch-based pyramid-HOG Class Weighting significantly outperforms the SMOTE in all metrics with mean differences being negative and substantial and corresponding p-values < 0.001. This indicates that the Class Weighting is the superior technique for handling class imbalance when using patch-based pyramid-HOG pipeline.

Compared to existing studies in Table 8. The proposed approach achieved a classification accuracy of 85.62%, outperforming the classification models applied to corneal imaging, such as the study by Al-Timemy et al. [17] and the study by Gandhi [19]. While Ahmed et al. [18] reported a higher accuracy of 98% using MobileNet-v2, their model was trained on over 16,000 augmented images, whereas the current study is done on a non-augmented dataset. This suggests that hand crafted features, when combined with optimized models, could match deep learning models in data-limited scenarios.

As shown in Table 9, all models maintained high generalization capability, with test accuracy above 82%. The Gaussian pyramid-HOG has achieved a test accuracy of 85.62%. To assess the reliability of this result, bootstrap resampling method is employed with 1000 samples, which yielded a mean accuracy of 85.64% (95% CI: 82.25%-88.76%) and a standard deviation of 0.0170. The relatively lower performance of Laplacian pyramid in both validation and test sets may stem from its emphasis on high-frequency details which, while important, may introduce noise.

Despite promising results, the proposed pipeline has limitations. First, the lower sensitivity in suspect class across all models indicates the inherent difficulty of detecting suspect classes. This challenge is consistent with clinical findings, where subclinical keratoconus often overlaps morphologically with normal corneas, Second, while the use of HOG feature enables interpretability, they may lack the expressiveness of learned deep features. For early diagnosis of the disease, the detection of suspect class from normal is crucial. Instead of relying solely on HOG features, attention networks like wavelet guided attention nets, multi-scale attention layers within CNN or a hybrid of CNN models and handcrafted features can be employed to detect the subtle variations in the corneal topography map of a suspect class. In addition, explainable AI, Grad cam visualization might help the model focus on the regions of the corneal maps that are highly discriminative in classifying suspect versus normal cases.