Microscopic image enhancement for accurate sickle cell disease identification using intuitionistic fuzzy driven retinex Authors: Jayasurya Ramakrishnan, Sriramakrishnan Pathmanaban, Vinodkumar Arumugam Published: 2026 ABSTRACT AbstractSickle cell disease (SCD) is a debilitating genetic blood disorder that must be identified and diagnosed early for effective treatment planning. Several factors hinder early identification in microscopic images, including uneven blood smears, lack of focus during capture, inconsistent illumination, and improper microscope settings. This paper addresses the aforementioned factors using the Intuitionistic Fuzzy-Driven Retinex (IFDR) method. The work contains two phases: microscopic image enhancement using IFDR and classification using DenseNet121. IFDR is a hybrid model that combines the Intuitionistic Fuzzy Index (IFI) with Single-Scale Retinex (SSR) to enhance uncertainty modeling and correct illumination variations in microscopic images. The IFI module enhances structural details and contrast in local area, while the SSR component performs global illumination balancing, resulting in diagnostically improved images. The enhanced images provide richer visual information, including clearer sickle cell edges, improved detection of overlapping cells, higher contrast, and less noise. For sickle cell identification, the enhanced images are used instead of the original images to train and test the DenseNet121 model, achieving competitive accuracy. The dataset used in this study was obtained from Kaggle. The performance of the proposed model is compared to the original and enhanced images. The classification accuracy for the original image is 0.877, whereas the proposed enhancement technique yields an accuracy of 0.964 for sickle cell identification. Comprehensive ablation studies were performed, confirming the superiority of the proposed hybrid method over its individual components. From experimental results, it is observed that the enhanced images yield better performance than the fine-tuned models and state of the art methods. Further, Grad-Cam-based explainable AI is employed to interpret the model classification results on enhanced images.Keywords: Intuitionistic Fuzzy, Image Enhancement, Deep Learning, Sickle Cell Disease, Single Scale RetinexSubject terms: Computational biology and bioinformatics, Diseases, Engineering, Mathematics and computing, Medical research Introduction The primary function of red blood cells (RBCs), which are essential components of human blood, is to transport oxygen from the lungs to organs1. RBCs typically have a 120-days lifespan, after which they are removed from circulation and replaced by freshly formed bone marrow cells. The RBC counts usually range from 4.7 to 6.1 million cells per micro liter for males, whereas 4.2 to 5.4 million cells per micro liter for females, depending on age, gender, and overall health2. The RBC analysis yields important insights into the physical state, which helps in diagnosis and treatment planning. Anemia is a hereditary disease caused by the production of abnormal hemoglobin, which leads to the formation of sickle-shaped RBC cells in blood3. In contrast, defective hemoglobin will have a crescent shape, commonly known as the “sickle” shape4. The main function of hemoglobin is to carry oxygen molecules across the body via blood; the sickle-shaped defective hemoglobin cannot carry oxygen, which leads to obstruction of blood circulation. The lifespan of a sickle cell is also notably smaller than that of a healthy cell; a defective cell’s lifespan stays up to 10 to 20 days, whereas a healthy cell can remain alive up to 120 days. Patients with sickle cell disease frequently suffer from intense pain, organ harm, anemia, organ damage and high susceptibility to infection5. Problem statements Early identification of sickle cell disease is extremely important for its treatment and prevention of problems6. Several factors can affect the quality of microscopic pictures, including uneven blood smear, lack of focus while capturing, inconsistent illumination, microscopic settings, and low resolution7. It can make it difficult to distinguish them from other cells or identify abnormalities, especially sickle cells, for proper diagnosis. Microscopic image enhancement is a crucial preprocessing step aimed at improving the visual quality and diagnostic utility of images captured through microscopes. These images, particularly in biomedical applications, often suffer from challenges such as low contrast, non-uniform illumination, noise, and blur due to limitations in sample preparation, staining variations, and optical systems. Several techniques were developed for image enhancement, including log transformation, gamma correction, histogram equalization, contrast-limited adaptive histogram equalization (CLAHE) and filtering methods8. These techniques are application-dependent and are particularly effective when applied to digital images. Limited research has focused on enhancing microscopic images through approaches such as fuzzy logic9 and Retinex-based methods10. In current decades, deep learning and artificial intelligence combo gives groundbreaking solutions across a wide range of real-world challenges, which include healthcare & medical imaging11,12, agriculture & food security13, industry & manufacturing14, transportation & smart cities15, security & surveillance16,17, and natural language processing (NLP)18. Advancements of artificial intelligence is now automated the diagnostic equipment with the help of deep learning on microscopic images to detect the sickle cells19,20. Using deep learning for the classification of results enhances it to a great extent and makes it more precise for the detection of sickle cells, enabling timely treatment and customized care plans21. Especially, DenseNet121 yields good classification accuracy on microscopic images for many applications like Fungi and Malaria Parasites detection22,23. There are several advantages to using DenseNet121 based classification in microscopic images. In DenseNet121, each layer receives input from all previous layers, which preserves fine-grained features. Enhance feature reuse, which is critical for recognizing small, subtle patterns like those in microscopic images. It reduces the vanishing gradient problem, which is especially helpful for detecting minute organisms or cells. DenseNet121 is more parameter-efficient compared with other CNNs with similar depth and has less overfitting on small medical datasets. These advantages are enhancing the performance of DenseNet121 on microscopic images. Related works The following literature was done with enhancement techniques on digital images using fuzzy logic, Retinex, etc. Ragavendirane and Dhanasekar propose a hybrid IFI-HER technique to improve low-light images by combining a new Intuitionistic Fuzzy Generator (IFG) with Histogram Equalization Retinex (HER)24. Driven by the detail and color loss in low light, the approach improves contrast with natural colors retained. The approach includes fuzzification of the image, creation of an IFI, application of HER, and choosing the optimum result using entropy optimization. The suggested method exceeds conventional methods in measures such as PSNR, AMBE, MI, and BRISQUE, with better visibility and edge precision. It performs very well in adverse illumination conditions. It is not as suitable for other types of degradations (such as motion blur) and real-time implementation because it has a high computational time. Babu et al. have introduced a new image enhancement method by combining Intuitionistic Fuzzy (IF) transformation with ABR-fractal fractional differentiation for enhanced global and local contrast in low-light images25. The method starts by transforming the image into an intuitionistic fuzzy space for enhanced brightness and reduction of ambiguity among pixel intensities. Then, a fractal-fractional differential operator under the ABR model is used to enhance local details and textures using a specially designed mask that relies on fractional order and fractal dimension. Lastly, a fusion mechanism integrates the globally and locally processed images into one output. This approach surpasses the classical enhancement methods in quantitative measurements such as entropy, contrast, and texture energy (through GLCM), and has greater edge and fine detail preservation. This renders it particularly efficient for low-illumination and low-contrast image situations. Yet, its reliance on multiple stages of transformation, with parameter adjustment, restricts real-time application and efficiency for distortions other than those of an illumination nature. Liu et al. have proposed a method to enhance low-light images, intended to recover details hidden by low illumination26. The algorithm is in a multi-stage process: image transformation from RGB to HSV color space, V component dynamic range stretching, estimation of illumination by adaptive gamma correction, and brightness enhancement using the Retinex model. The approach is computationally effective, surpassing existing techniques in qualitative as well as quantitative evaluation. RBFA captures higher visibility, enhanced brightness uniformity, and natural color retention and is the best for low-light conditions. The following literature is utilizing deep-learning models for sickle cell identification without image enhancement27. The most commonly used dataset for this classification of sickle cells is the ErythrocytesIDB dataset, while most of the work is done on segmentation and a few work on classification. Visual geometric group (VGG) is a model proposed by Garcia-Garcia in a review of the deep learning model. They proposed a model VGG16 with an accuracy of 0.927, which then paved the way for VGG19 an accuracy of 0.9728. ReNet architecture is RNNs with spatio-temporal data dimension29. In ReNet, the number of RNNs increases linearly for each layer based on the input dimension, and the accuracy it has is 0.68 in SCD classification. Deep learning models like Deep CNN are used to build a CNN model from scratch with the convolution layer, pooling layer and fully connected layer30. It is used in the detection of SCD, which gives us a classification accuracy of 0.81. Table 1 shows the recent findings on SCD, which were collected from datasets. Table 1The summary of recent findings on SCD from different data sources.Author DetailsMetric NameModel(s) Name and FindingsObserved Key LimitationsPatel et al., (2026)31AccuracyResNet-50 – 0.930a. Limited dataset size and diversityb. Class imbalance issuesc. Transfer learning domain gapDenseNet-121 – 0.915EfficientNet-B0 – 0.900Malarselvi et al., (2025)32AccuracyLogistic Regression – 0.923a. Morphological overlap with sickle cellsb. Annotation variabilityc. Image variations in illumination, focus, staining, and noiseDeo et al., (2024)33AccuracyCNN + LSTM – 0.926a. CNN-LSTM architectural mismatchb. High computational complexityc. Dependence on image qualityCNN InceptionV3 – 0.901Extreme Learning Machine – 0.876Petrovic et al., (2025)34F1 ScoreRandom Forest and Extra Trees – 0.9071a. Complex modelb. Dependency on base learner diversityc. Overfitting risk with ensemblesGradient Boosting – 0.8732Carrasco et al., (2025)35F1 ScoreResNet-50 – 0.88a. Size of the annotated data setb. Images variabilityc. Overfitting riskEfficientNet-b0 – 0.83 The summary of recent findings on SCD from different data sources. a. Limited dataset size and diversity b. Class imbalance issues c. Transfer learning domain gap a. Morphological overlap with sickle cells b. Annotation variability c. Image variations in illumination, focus, staining, and noise a. CNN-LSTM architectural mismatch b. High computational complexity c. Dependence on image quality a. Complex model b. Dependency on base learner diversity c. Overfitting risk with ensembles a. Size of the annotated data set b. Images variability c. Overfitting risk Research gaps and limitations of existing methods From the literature, despite significant advancements in microscopic image enhancement and sickle cell detection, several limitations persist in existing approaches. Most conventional methods primarily focus on either contrast enhancement or illumination correction independently. It leads to suboptimal performance when dealing with complex microscopic images that exhibit both uneven illumination and inherent uncertainty. Additionally, many techniques fail to preserve fine structural details, such as cell boundaries and overlapping regions, which are critical for accurate diagnosis. Traditional methods like histogram equalization, CLAHE, or Retinex variants end up over-enhancing images or losing important, subtle diagnostic information in medical images. Though fuzzy logic–based methods have been promising, most methods so far have used only traditional fuzzy sets, which only address one level of uncertainty (membership). This reduces the method’s capability to address the vagueness and ambiguity inherent in medical images, especially in noisy or low-light situations. Despite considerable progress in general image enhancement techniques, microscopic image enhancement for sickle cell detection remains underexplored and underdeveloped. While a few studies have attempted to apply fuzzy logic for microscopic enhancement, they often overlook the uncertainty and ambiguity in pixel intensity, especially in borderline or partially deformed cells. Motivations These conventional approaches also struggle with low contrast and uneven illumination inherent in stained blood smear images. Other factors are high noise levels and cell overlap that obscure critical diagnostic details, and inadequate edge preservation, which is essential for detecting the elongated or crescent shapes of Sickled cells. Existing methods lack robustness and adaptability to variations in staining protocols, microscope settings, and magnification scales, which affects their generalizability. Furthermore, existing deep learning-based classification models often rely on raw or insufficiently enhanced images, thereby limiting their ability to achieve robust and reliable performance. Despite advancements in image processing, there remains a notable gap and increasing need for specialized image enhancement techniques for microscopic images, especially those used in the identification of sickle cell disease. These gaps and challenges highlight the need for a computationally efficient hybrid framework that jointly performs local and global enhancement, uncertainty modeling, and illumination correction to improve microscopic image quality for accurate sickle cell identification. Novelty of the proposed work To address the above mentioned challenges, this study proposes a hybrid IFDR model that integrates IFI for effective uncertainty representation and structural enhancement with SSR for global illumination compensation. This combined approach aims to produce diagnostically meaningful enhanced images, thereby improving the accuracy and reliability of sickle cell identification. The novelty of the IFDR framework lies in: The integration of IFI and SSR to jointly perform uncertainty modeling, local structural enhancement, and global illumination correction. The integration of IFI and SSR to jointly perform uncertainty modeling, local structural enhancement, and global illumination correction. The ability to produce diagnostically meaningful enhanced images with improved edge preservation, reduced noise, and better contrast difference in cell overlapping. The ability to produce diagnostically meaningful enhanced images with improved edge preservation, reduced noise, and better contrast difference in cell overlapping. The use of enhanced images to improve downstream classification performance (DenseNet121), demonstrates practical effectiveness. The use of enhanced images to improve downstream classification performance (DenseNet121), demonstrates practical effectiveness. Difference from existing fuzzy-based and retinex-based methods In contrast to conventional fuzzy-based methods that mainly emphasize local contrast enhancement, the proposed approach employs the IFI to more effectively capture uncertainty while maintaining structural integrity. Similarly, while conventional Retinex-based methods focus on illumination correction, they often neglect structural enhancement and may introduce artifacts. The proposed framework differs by combining both approaches in a unified manner, where IFI enhances local structures, and SSR performs global illumination balancing. Materials and metrics This section discusses the details of the dataset used. Further, the subsection confers the validation metrics of the classification results. Materials used The dataset includes images gathered from the Teso area in Uganda, particularly from the districts of Kumi and Soroti. The dataset is an open-access compilation found on Kaggle36,37, featuring labeled cell images that indicate whether sickle cells are positive or negative. Dataset images are depicted in Fig. 1, which includes sickle cell positive (a) and negative (b). Sickle cells are highlighted with rectangular boxes in Fig. 1(a). The dataset is generated from 140 patients provided with their blood samples, and extracted the features using two staining techniques the field and leishman stains. The samples are then photographed under a microscope and used for evaluation in two classes. Sickle cell positive consists of 422 images and 147 negative images. Sickle cell negative consists of clear cells, which are extracted from healthy patients having images of regular healthy RBC. To solve the class imbalance issue, inverse frequency class weighting was used during the training process38. Sample images from dataset. (a) Sickle cell positive, (b) Sickle cell negative. Assessment metrics To assess the IFDR performance, enhancement and classification qualitative metrics were used39. To find the similarity between original and enhanced images, peak signal to noise ratio (PSNR), mean square error (MSE), structural similarity index measure (SSIM) and correlation coefficient (CC) were used40,41. PSNR measure the quality of reconstruction between the original and enhanced images. MSE is a metric used to measure the pixel intensity difference between an original image and its enhanced image. Both metrics can be calculated using the below equations. I – original image, E – enhanced image, MAX – maximum pixel value of the image, and –image dimension. SSIM is a widely used metric to evaluate visual similarity between the original and enhanced image rather than just pixel-wise differences. It measures the visual similarity such as luminance, contrast and structures. The value of SSIM is fall between − 1 and 1; one for identical images and zero for no similarity. It can be computed using Eq. (3). Where, – mean intensity of original and enhanced image, , – variance, – covariance, and , – constants to stabilize division. Correlation coefficient (CC) is a statistical measure used to evaluate the degree of similarity or linear relationship between the original and enhanced images. Here, Cov(X, Y) – Covariance between original image (I) and enhanced image (E), (I), D(E) – variance of I and E. The classification metrics are derived using a confusion matrix table. It offers insights into the model’s effectiveness by showing the count of true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN) and is explained in Table 2. This matrix aids in assessing the trade-off between precision and recall and is beneficial for identifying misclassification errors. It is especially advantageous when managing imbalanced datasets, as it provides a better insight into the model’s advantages and limitations beyond mere accuracy. Table 2The confusion matrix for the sickle cell identification.Predicted ResultsSickle Cell PositiveSickle Cell NegativeActual ResultsSickle Cell PositiveTPFNSickle Cell NegativeFPTN The confusion matrix for the sickle cell identification. For quantitative assessment, several classification metrics are employed for a comprehensive evaluation of the DenseNet121 model’s performance. The accuracy of the model’s performance can be computed as follows. Precision refers to the ratio of accurately identified positive cases to the total predicted positive cases. Recall (Sensitivity) assesses the capability to accurately recognize all true positive instances. The F1-Score represents the harmonic mean of precision and recall, offering a balanced assessment of both. . IFDR Methodology The block diagram of the proposed IFDR is shown in the Fig. 2. The proposed work contains two phases: image enhancement and image classification. Image enhancement is performed using an Intuitionistic Fuzzy-based Retinex mathematical model, while classification is carried out using the DenseNet121 architecture. The block diagram shows a series of image preprocessing techniques motivated by the enhancement of the original image and gives a better analysis of the classification problem. Initially, the original image is taken from the data source and then enhanced using the proposed Intuitionistic fuzzy-based Retinex enhancement. The enhanced images are subsequently fed into the same DenseNet121 model to assess the effect of preprocessing on the accuracy of classification. The study seeks to show the effectiveness of image enhancement in increasing model accuracy and robustness by comparing results from original and enhanced images. The block diagram of the proposed IFDR. Phase I: image enhancement - intuitionistic fuzzy driven retinex (IFDR) Under the enhancement approach, intuitionistic fuzzy is used to increase the light and brightness of low-contrast microscopic images43. As a factor of application, the image sets are microscopic images extracted from blood samples, where images are low-light and lack features for analysis. Low-light image enhancement is a crucial function in micro imaging, where some key details are usually lost because of insufficient light or illumination in the machine settings or sensor constraints. The basic idea of intuitionistic fuzzy is to enhance the visibility of image features and make it visually good44. Enhancement procedures are not just necessary to enhance overall image quality but also to highlight low-contrast features that can be clinically important. In such applications, the application of Intuitionistic Fuzzy Set Theory appears to be a promising avenue. As compared to conventional fuzzy sets that model only one kind of uncertainty through membership functions, intuitionistic fuzzy sets include membership and non-membership degrees and a hesitancy component for supporting richer modeling of uncertainty. This phase explains the hybridization of Intuitionistic fuzzy and single-scale Retinex for microscopic image enhancement. Intuitionistic Fuzzy Logic is an extension of traditional fuzzy logic introduced by Atanassov42. Unlike classical fuzzy sets, which define each element with a single membership value (µ), intuitionistic fuzzy sets define three parameters for each element: Membership degree (µ), non-membership degree (ν), and Hesitation degree (π) and Intuitionistic Fuzzy Index (IFI). With Contrast Limited Adaptive Histogram Equalization (CLAHE)45 included, the technique guarantees localized contrast enhancement, efficiently processing low-contrast as well as unevenly lit images. The implementation contains nine steps, and the algorithm is given below and step-by-step explanation is as follows. Intuitionistic Fuzzy Driven Retinex (IFDR). Step 1: Image Fuzzification. The process of enhancement starts with fuzzification, in which intensity values of an obtained image are mapped to a value between 0 and 1. This guarantees that the input image is projected into a space where subsequent transformation can be conducted. This process scales the pixel values of each color channel (Red, Green, Blue) to the range [0, 1], converting the image into a fuzzy domain representation as per Eq. (9). Where, is the input pixel intensity, ; M and N denote the number of rows and columns respectively, and are the minimum and maximum intensity values of the input image, respectively, and epsilon (ϵ) is a small constant added to avoid division by zero. Step 2: Membership (µ) and Non-membership (γ) Function Computation. After being fuzzified, membership (µ) and non-membership (γ) values are determined based on a parameter , which determines the extent of enhancement. Where, is the fuzzified intensity and in our experiments. Step 3: Selection of optimal value. These functions (Eq. (10) and Eq. (11)) depend on a tunable parameter that controls the steepness and sensitivity of the transformation. The -value plays an important role in the calculation of the hesitancy function in the proposed enhancement technique. It acts as a tuning parameter that directly depends on the calculation of membership and non-membership values, thereby affecting the Intuitionistic Fuzzy Index (IFI) and ultimately the quality of the enhanced image. Selecting the optimal -value is crucial to ensure that the enhancement process produces visually meaningful and information-rich outputs. In scenarios where a limited number of images are available, and enhancement quality is the primary objective, a systematic optimization approach can be used to determine the most suitable -value. The range of -values would be from 0.1 to 1.0 in general. Each -value represents a different degree of influence that leads to various features enhanced in the enhancement process. For each -value from the range, the entropy of the resulting enhanced image is calculated. The entropy for the enhanced image at each channel (c) is computed using Eq. (12). Here, represents the probability of intensity level in channel . The total entropy of an image at each is: The goal is to identify the optimize k-value that yields the maximum entropy using Eq. (14), as this indicates the highest potential for information content and contrast in the enhanced output. Step 4: Hesitancy Degree Estimation. The hesitancy value is representing the uncertainty between the degree of membership and non-membership. It shows how much an element belongs to a class versus how much it does not. This parameter encodes the intuitionistic component, which contributes significantly to the overall improvement. The hesitancy degree represents the vagueness related to a pixel’s fuzzy representation, and it is calculated as: Step 5: Intuitionistic Fuzzy Index (IFI) Computation. The IFI is subsequently determined through the addition of the membership function and the hesitancy of the image. This composite index captures confidence as well as uncertainty in the image structure, enhancing low-level features in bright and dark areas. It is calculated below. Step 6: CLAHE Enhancement of IFI. The output IFI is applied with CLAHE to amplify local contrast. CLAHE is further applied to refine the results by transforming the image into IFI form and adapting contrast improvement for various regions. It is most suitable for medical and microscopic imaging, where tiny details should be emphasized without over-boosting noise. It works on small tiles and varies contrast enhancement to avoid noise over-amplification. CLAHE divides the given image into non-overlapping tiles of size The parameters such as Clip_Limit = 2.0, Tile_Grid_Size = 8 × 8 were chosen based on commonly adopted ranges in the literatures. The number of tiles is computed as: Compute the histogram for each tile : Where, is the gray level. Apply contract limiting (C), which is the maximum number of pixels allowed per bin; if , then extra pixels . To compute total excess (E): Here, is the local histogram value of intensity , C is the constant threshold, and L is the total number of intensity levels. Redistribute the total excess E evenly among all bins: Here, is the clipped and redistributed histogram. Compute the cumulative distribution function (CDF) for : Then, the mapping function for each intensity becomes: To avoid artifacts, interpolation between tiles would be: Where, is the input image intensity at position , is the transformation function of the ith neighboring tile, is the interpolation weight corresponding to the proximity of the ith tile to the pixel , sum is taken over the 4 nearest tiles surrounding the pixel, enabling smooth transitions between tiles. Step 7: Image Defuzzification. The final step is defuzzification, wherein the enhanced image is re-projected onto the original intensity scale with the introduction of hesitancy component adjustments to preserve structure consistency. Lastly, the image is defuzzified by rescaling the amplified IFI to the original intensity range with hesitancy as a subtractive factor: Where, and are the original image channel’s minimum and maximum values. For color images, the operation is performed independently on each of the channels (red, green, and blue) to ensure that improvement is carried out equally without affecting the color balance. The approach efficiently handles the uncertainty, which includes low contrast, noise and blurred cell edges and enhances the brightness as well as representation of details based on local features, image contrast, hence ideal for microscopic image analysis like sickle cell anemia classification in blood smear images. Through the dynamic contrast adjustment based on fuzzy logic methodology, this method provides a strong alternative to conventional enhancement techniques that promise better image quality for subsequent analysis. Step 8: Logarithmic transformation. The Single-Scale Retinex (SSR) is an image enhancement method modelled after the human visual sighting system46. The main idea is to increase the visual quality of photographs by correcting lighting changes (reflectance and illumination) and boosting the contrast of the image. The center/surround Retinex hypothesis, originally proposed by Jobson et al., explains how the human visual system perceives color and light under varying illumination conditions47. The following steps are applied on the enhanced image . Although contrast correction is performed in the earlier steps, further enhancement is required through light correction by separating reflectance and illumination components. Using a Gaussian-based surround function to estimate the illumination and reflectance component of the image and applying a logarithmic adjustment to the input image (is the basic idea of SSR. The reflectance is extracted by subtracting the illumination estimate from the logarithmic-transformed image after this surround function has been convolved with the original image. With less variance in intensity and better local contrast, the reflectance component functions as the improved image. Mathematically, SSR is formulated as: Where represents the input image distribution over the ith spectral band, ) is the Gaussian surround function responsible for illumination estimation which is computed using Eq. (26), and ‘*’ denotes the convolution operation and denotes Retinex output image after log transformation. In Eq. (26), was chosen through grid search method which has ten candidate set of Method can apply SSR and compute the enhancement metrics using Eq. (1) to Eq. (4) to select the optimal for the proposed implementation. Inverse-square, Gaussian, and exponential models of the surround function are frequently used; each has a distinct impact on the enhancing result. The SSR has several disadvantages despite its ability to reduce noise and enhance local details and contrast. The technique is not suited for images with extremely non-uniform lighting since it cannot compress the dynamic range and reproduce tone simultaneously. In addition, because of the abrupt lighting correction, SSR produces halo artifacts at the edges of objects. Despite these problems, SSR remains a standard entry point in computer vision and medical imaging, particularly for rendering low-contrast images visible. Since it is computationally lightweight, the technique may be applied in real-time situations where speed is crucial. To overcome its limitations, researchers have attempted more sophisticated varieties such as Multi-Scale Retinex (MSR) and MSR color restoration (MSRCR). It applies various surround functions at various scales to enhance color uniformity and dynamic range compression. SSR plays a major role in enhancing blood cell contours in sickle cell classification to make the automated detection models detect them more easily. Step 9: Clipping and normalization. A histogram is computed for each channel individually, and the pixel intensity values are traversed to determine the low and high threshold values for clipping. All pixel values below the low threshold are set to the low threshold, and those above the high threshold are set to the high threshold. Subsequently, the image is normalized back to the 0–255 range using min-max normalization to enhance contrast while preserving image structure. Convert the image back to uint8 format for the enhanced image . Phase II: Image Classification In recent times, deep learning has emerged as a strong method for automated diagnosis of diseases. For sickle cell classification, DenseNet121 is employed with the enhanced image , given by the IFDR. DenseNet121 is a CNN architecture and is able to extract features from given images. It can learn more deeply about the sickle cell structure, shape, position, and edges if trained with an IFDR-based enhanced image. This phase is aimed at enhancing gradient flow and parameter efficiency using dense connectivity to enhance images. Each layer of DenseNet121 takes inputs from all its preceding layers and transfers its feature maps to all the following layers. Such a pattern of connectivity assists in reducing the vanishing gradient problem, encourages feature reuse, and decreases the number of parameters as compared to conventional deep networks. Our preferred model for sickle cell classification effectively differentiates sickle cells and normal cells accurately using enhanced images. The model is trained to identify required features from microscopic images and classify them efficiently through a multi-stage feature extraction and classification process. DenseNet121 is made up of four densely connected blocks, which are interspersed by transition layers comprising batch normalization, convolution, and pooling processes48. The model architecture for DenseNet121 is shown in Fig. 3. The core structure of our model is a DenseNet121 architecture that is based on dense connections between layers to provide feature extraction and gradient flow. The model takes as input 128 × 128 × 3 RGB images, which are initially passed through an initial convolutional and pooling layer for low-level feature extraction. The key components of the DenseNet121 model are explained in the sections below. The DenseNet121 architecture for sickle cell classification. The input layer is initially zero-padded to provide spatial consistency. A convolutional layer (64 filters, kernel size 7 × 7, and stride 2) is utilized, and then batch normalization is followed by a ReLU activation function. A max-pooling layer (kernel size 3 × 3, and stride 2) is employed to reduce the data in feature maps to reduce computational complexity without losing vital spatial information. Each Dense block contains 1 × 1 convolutions for reducing dimensionality and 3 × 3 convolutions for extracting spatial features. Batch normalization and ReLU activation to stabilize training and add non-linearity. Concatenation layers are used to preserve information flow between layers. This method ensures maximum feature reuse, enhancing efficiency and accuracy. After passing through the 3 dense blocks, three separate Global Average Pooling (GAP) layers are applied at different stages after each major dense block. GAP layers significantly reduce the number of parameters while retaining critical spatial information, making the model lightweight and efficient. Three dense layers, each with 64 neurons, are applied with batch normalization and dropout regularization to prevent overfitting. These dense layers are combined into one feature representation (192 neurons total) before classification. A 16-neuron dense layer fine-tunes the extracted features before classification. The model has 7.2 million trainable parameters, well-organized to optimize feature extraction while being computationally feasible. Batch normalization and dropout regularization are utilized at various stages to enhance generalization and prevent overfitting. The model is trained with a binary cross-entropy loss function, optimized with an Adam optimizer for rapid convergence49,50. The last output layer (2 neurons, Softmax activation) gives the probability of the input image being in either the normal or sickle cell category. The enhanced images are used for classification (). The images are resized to 128 × 128 pixels and normalized to ensure model input uniformity. An 80 − 20 split is used for training and testing the model with a uniform distribution of classes maintained to handle the data bias at the image level. Class weights are calculated dynamically using the inverse frequency class weighting method based on training label distribution to handle class imbalance and achieve fair learning between categories. To ensure class balance between sickle and normal cells, a stratified sampling strategy was employed during the splitting process. Results and discussion The experiment was developed using Python 3.11 with the necessary packages. The following subsection discusses the qualitative and quantitative results of image enhancement and classification. Qualitative results of proposed IFDR-based image enhancement Qualitative results of the source and proposed methods for sickle cell image enhancement are given in Table 3. Table 3 results depict the annotated image enhancement results on the dataset, which has both sickle-positive and negative cells. In Table 3, rows II and III show the results of conventional methods, and Row IV depicts the proposed results. Proposed image enhancement results are shown in row IV. From the qualitative annotated results, it can be seen that the proposed IFDR-based enhanced images are effective; each sickle cell has clear boundaries. It enhances cell boundaries and preserves cellular morphology, enabling the identification of a greater number of sickle cells compared to the original images. It can adjust the contrast, sharpness, light, and illumination for each image locally. IFDR-based image enhancement gives more detail. From the annotated results, the proposed IFDR method enables the identification of a greater number of sickle cells with more clearly defined boundaries, demonstrating improved visual and qualitative performance. These enhanced images lead the way for deep learning models to learn well with the sickle cell features map and produce increased classification accuracy for sickle cell identification than the image without enhancement. Table 3Qualitative results of IFDR-based image enhancement. Qualitative results of IFDR-based image enhancement. Quantitative results of proposed IFDR-based image enhancement The objective of this work is to improve the accuracy of sickle cell detection through image enhancement. While the dataset provides ground truth for classification, it does not include ground truth for enhancement-based validation. Although, quantitative validation metrics were used to assess the similarity between original and enhanced images using PSNR, SSIM, and MSE and the results are given in Table 4. The statistical results show the model enriched the significant changes from the original image through adjusts the contrast, sharpness, light, and illumination locally. Table 4Quantitative results of IFDR-based image enhancement.Category of ClassPSNR(dB)MSESSIMCorrelation CoefficientPositive13.4723562.1580.59290.8559Negative13.7482983.990.57510.8892 Quantitative results of IFDR-based image enhancement. PSNR values are relatively low, indicating that the enhanced images differ noticeably from the originals. This suggests that the enhancement process introduces substantial intensity modifications, which is expected when contrast or structural features are amplified for better visibility of sickle-shaped cells. High MSE values further confirm significant pixel-level changes between raw and enhanced images. Negative class shows comparatively lower MSE, implying that enhancement preserves background/normal cell structures more consistently than pathological regions. SSIM values indicate moderate structural preservation. Although enhancement alters pixel intensities, the essential structural information (cell boundaries and morphology) is reasonably maintained. Slightly higher SSIM in the Dataset (Positive: 0.5929) suggests better preservation of structural patterns relevant for identifying sickle cells. The correlation coefficient shows that high values indicate a strong linear relationship between the original and enhanced images. The dataset shows higher CC, especially for the negative class (0.8892), meaning enhancement does not distort normal cell patterns significantly. Sickle cell classification results with component ablation To evaluate the effectiveness of image enhancement techniques in sickle cell classification, and to ensure the proposed model stability, the component ablation study involved the following four components, which were created from the dataset. i.Baseline – Original image without image enhancement and DenseNet121. Baseline – Original image without image enhancement and DenseNet121. ii.Intuitionistic Fuzzy Index (IFI) – IFI enhancement with DenseNet121 classification. Intuitionistic Fuzzy Index (IFI) – IFI enhancement with DenseNet121 classification. iii.Single Scale Retinex (SSR) – SSR enhancement with DenseNet121 classification. Single Scale Retinex (SSR) – SSR enhancement with DenseNet121 classification. iv.Proposed IFDR– IFDR enhancement with DenseNet121 classification. Proposed IFDR– IFDR enhancement with DenseNet121 classification. The model’s performance in sickle cell classification was assessed using qualitative and quantitative metrics. The following sections contain the component ablation study, which arranges the results in the given order. The training and testing ratios are 80 and 20, respectively. The final results were computed using 5-folds cross-validation method with 20 epochs. Performance analysis of baseline component In this experiment, the DenseNet121 model was trained on unprocessed or original images. The performance metrics obtained from the original dataset serve as a baseline for comparison. Among 114 testing images, the model can predict 100 images correctly and misses 14 images as FP and FN. This misclassification happens due to the poor illumination of the microscopic setting. The model is unable to differentiate the sickle cell patterns from normal cells for those 14 images. Further, the quantitative fold -wise result for the original dataset is shown in Table 5. The baseline performance was achieved by training the model directly on the original images. It shows that the model yields an accuracy of 0.87 on the original images. The validation accuracy and loss curves indicate how well the model converges during training. There is a certain dip in accuracy on the 11th epoch, which means the model is memorizing the training data but failing to generalize. The ROC curve for this dataset illustrates its ability to distinguish between sickle cell-positive and healthy samples, and the AUC value for the original dataset is 0.91, demonstrating a fairly good classification capability. Nonetheless, the performance was constrained because of the inherent noise and discrepancies in medical imaging. Table 5Quantitative results on baselines for sickle cell classification using DenseNet121.Fold NumberAccuracyPrecisionRecallF1-ScoreAUCFold 10.87120.82650.85430.83910.9054Fold 20.88350.83890.86670.84980.9176Fold 30.87940.83470.86190.84530.9135Fold 40.87480.83010.85820.84200.9098Fold 50.87710.83370.86310.84840.9135Mean0.87710.83270.86080.84490.9119 Quantitative results on baselines for sickle cell classification using DenseNet121. Performance analysis of IFI component The dataset was created with the enhanced images using the Intuitionistic Fuzzy method. After applying Intuitionistic Fuzzy enhancement, the DenseNet121 classifier was trained on the dataset to observe the impact of enhancement and improved feature extraction. These results demonstrate whether the fuzzy logic-based method contributes to improved classification. Table 6 show that results of Intuitionistic Fuzzy image enhancement, which plays a significant role in the sickle cell classification. On the raw images, the accuracy was 0.87, and Intuitionistic Fuzzy enhancement yields an accuracy of 0.90 for cell classification. From the obtained results, enhanced images contain clearer sickle cell edges, textures, and contrast, which help DenseNet121 extract more meaningful and distinct features at each layer. A higher AUC score (0.92) in the ROC curve would indicate better discrimination between classes in enhanced images. From the results mentioned in this section, the fuzzy logic-based enhancement contributes to the classification accuracy of the conventional method, but is not enough to achieve the goal. Table 6Quantitative results on intuitionistic fuzzy enhancement for sickle cell classification using DenseNet121.Fold NumberAccuracyPrecisionRecallF1-ScoreAUCFold 10.89820.90140.82650.85530.9178Fold 20.90760.90980.83890.86570.9265Fold 30.90410.90670.83240.86140.9236Fold 40.90170.90390.83070.85920.9214Fold 50.90590.90640.83680.86380.9253Mean0.90350.90560.83300.86100.9229 Quantitative results on intuitionistic fuzzy enhancement for sickle cell classification using DenseNet121. Performance analysis of SSR component The SSR enhancement was applied to the original images to correct illumination inconsistencies and improve local contrast before training the DenseNet121 classifier. This method aims to enhance the fine details in RBC structures, potentially improving classification accuracy. SSR enhancement bettered performance by standardizing lighting conditions and boosting local contrast. From Table 7, SSR-enhanced images yield higher accuracy with dense layers than fuzzy and original images. Hence, SSR-based enhanced images boost the quality of initial features, making the reused information richer and more informative throughout the network. Standard or low-quality images may cause the network to lose important details in early layers. Table 7Quantitative results on SSR Enhanced Images for sickle cell classification using DenseNet121.Fold NumberAccuracyPrecisionRecallF1-ScoreAUCFold 10.93420.92870.89760.91150.9634Fold 20.94160.93680.90590.91970.9708Fold 30.93910.93350.90180.91640.9681Fold 40.93680.93120.89940.91430.9665Fold 50.94130.93540.90540.91880.9710 Quantitative results on SSR Enhanced Images for sickle cell classification using DenseNet121. Performance analysis of proposed IFDR Finally, the model was trained on enhanced images using the proposed IFDR. This experiment evaluates whether the IFDR yields superior results to the conventional methods. From Fig. 4; Table 8, the total number of misclassified images is four, which helps to achieve a high accuracy of 0.964 for sickle cell disease identification using DenseNet121. Here, image enhancement ensures the critical details of sickle cells are preserved, so even deep layers retain contextually important patterns. High-quality inputs mean more stable and discriminative gradients during backpropagation. This leads to faster convergence and better optimization in DenseNet with many skip connections. Enhanced images reduce noise and irrelevant details, which DenseNet121 might otherwise memorize. This leads to better generalization and lower overfitting. The qualitative results of proposed IFDR. (a) Confusion matrix, (b) Accuracy curve, (c) Loss curve, (d) ROC curve. Table 8Quantitative results on proposed IFDR-enhancement for sickle cell classification using DenseNet121.Fold NumberAccuracyPrecisionRecallF1-ScoreAUCFold 10.96050.97340.92610.94720.9598Fold 20.96780.98010.93480.95490.9685Fold 30.96430.97780.93150.95170.9641Fold 40.96290.97620.92940.95010.9627Fold 50.96890.98010.93340.95340.9664Mean0.96490.97750.93100.95140.9643 Quantitative results on proposed IFDR-enhancement for sickle cell classification using DenseNet121. The enhancement in performance indicates that utilizing various improvement methods enables the model to identify more significant features, resulting in superior classification. The optimal classification outcomes were attained through a hybrid enhancement method, where the IFI and SSR techniques were utilized to preprocess images before training. This method utilized the benefits of contrast enhancement (Fuzzy) and illumination normalization (SSR), resulting in a peak accuracy of 0.9649 and an AUC score of 0.9643. The effectiveness of each technique was evaluated using important assessment metrics, and the findings are summarized in Table 9. From the table, it can be observed that the accuracy of the classification model is gradually increasing with our proposed methods and gives a reasonable improvement in accuracy from the original image. The results show that the image preprocessing enhancement technique can significantly impact the performance of sickle cell classification results. The improvements suggest that preprocessing helps enhance necessary features in the image data, making classification more robust and accurate. Table 9Overall comparison of sickle cell disease identification results.Name of the ComponentAccuracyPrecisionRecallF1-ScoreAUC ScoreBaseline0.8771930.8327720.8608520.8465790.911968IFI Component0.9035090.9056320.8330630.8678330.922921SSR Component0.9385960.9331290.9020280.9173150.967951Proposed IFDR0.9649120.9775280.9310340.9537150. 964300 Overall comparison of sickle cell disease identification results. From Table 9, the proposed method demonstrates improved performance across most evaluation metrics, while SSR shows slightly higher AUC. The slight variation in AUC may be attributed to the trade-off between sensitivity and specificity, whereas IFDR provides more stable and reliable performance across different evaluation criteria. However, the performance of a model should be evaluated comprehensively across multiple metrics rather than relying on a single measure. While SSR shows a marginal advantage in AUC, the proposed IFDR method consistently demonstrates superior or more balanced performance across other key metrics, such as accuracy, precision, recall, and F1-score. The proposed IFDR framework is designed to jointly optimize structural enhancement (via IFI) and illumination correction (via SSR), resulting in improved overall image quality and better generalization for classification. The paired t-test is selected for performance comparison, which was performed on the same cross-validation folds for all methods to analyze the statistical significance. A paired t-test was conducted between the proposed and baselines components to prove the statistical significance, which is given in Table 10. It was calculated from the difference in the accuracy results. The obtained p-value (< 0.05) indicates that the improvement achieved by the proposed IFDR model is statistically significant from the baseline accuracy, such as IFI and SSR. A high paired t-value occurs when the proposed method consistently outperforms the comparator across folds with very low variation in fold-wise differences. Table 10The pairwise statistical significance tests between proposed and baseline accuracies.Name of the Modelt-score on IFDRp-value on IFDRStatistical SignificanceBaseline62.840.00000038YesIFI Component109.460.00000004YesSSR Component68.450.00000027Yes The pairwise statistical significance tests between proposed and baseline accuracies. Quantitative performance comparison of the pre-trained model on SCD detection is given in Table 11. The existing results were obtained by using a pre-trained model for feature extraction, followed by classification using Bayes or SVM classifiers56,35. Existing models are extracting features from original images, which often suffer from low contrast, blur, or artifacts. Without image enhancement, useful structures may remain hidden, leading to poor feature extraction. Therefore, without proper enhancement, existing methods yielded poor performance in SCD detection. While comparing, the proposed enhanced images gave the proper shape of sickle cells and thus improved the accuracy of SCD detection. Table 12 shows the SCD detection results by existing methods which use various techniques. It shows that the proposed IFDR-based image enhancement outperforms the existing methods. Table 11Classification performance comparison of pertained models on SCD detection.Name of the ModelAccuracyPrecisionRecallF1-ScoreDenseNet - Bayes0.82600.83470.82670.8282DenseNet - SVM0.90590.91230.90660.9061ResNet - Bayes0.59100.61300.59730.5291ResNet - SVM0.68050.70260.67880.6469MobileNet - Bayes0.85300.87480.85300.8539MobileNet - SVM0.91210.91420.91280.9107ResNet500.94900.95000.94750.9350GoogLeNet0.85710.89000.74500.7800ResNet180.89800.94000.81000.8500Proposed IFDR + DenseNet1210.96490.97750.93100.9537 Classification performance comparison of pertained models on SCD detection. Table 12Classification performance comparison with existing methods on SCD detection.Name of the methodAccuracyLogistic Regression [Malarselvi et al., 2025]320.9230Hyperparameter Optimization [da Silva et al. 2020]510.9254CNN + LSTM [Deo et al., 2024]330.9260Contrastive learning [Patel et al., 2026]310.9300Bag-of-Visual-Words [de Faria et al. 2018]520.9367Morphological analysis [Rodrigues et al. 2016]530.9459Integral-Geometry [Gual-Arnau et al. 2015]540.9610IFDR Enhanced Images with DenseNet1210.9649 Classification performance comparison with existing methods on SCD detection. Model Interpretation and Post-Processing For qualitative analysis, we employed Grad-CAM from Explainable AI (XAI)55, which shows the areas the model pays attention to at the time of classification. The ground-truth region annotations are not available for objective validation of the highlighted areas. Therefore, the Grad-CAM analysis is qualitative in nature and is intended only to provide visual insights into the model’s attention patterns, rather than to claim definitive interpretability or quantitative localization performance. The visualizations effectively show that the model trained using augmented images demonstrates improved feature localization and attention as opposed to original images, substantiating the influence of our preprocessing strategy. It allows DenseNet121 to focus on more meaningful sickle cell regions, improving interpretability and classification accuracy. Table 13 shows the DenseNet121 interpretation results given by the Grad-CAM on both classes with and without IFDR-based image enhancement. These images are generated between the convolution layers, one of which retains the sickle cells spatial information. Feature maps from these layers preserve where sickle cells are present in the image. From Table 13, red, yellow, and blue colors indicate high, medium and low feature spatial importance, respectively, while classification. Table 13The XAI-based DenseNet121 model interpretation using Grad-CAM with and without enhancement. The XAI-based DenseNet121 model interpretation using Grad-CAM with and without enhancement. On cautious inspection of the Grad-CAM visualizations, it appears that the model might be focusing on red areas with dense cellular boundaries, elongated contours, and irregularly shaped erythrocyte structures. Highlighted regions appear to correspond with some of the crescent-shaped or curved cell morphologies that are often associated with sickle cell disease. This could suggest that the network is learning shape features that are useful for positive-class discrimination. However, these observations should be taken with a grain of salt, as Grad-CAM yields coarse attention maps but not exact cellular localization. The highlighted areas may also correspond to overlapping cells, staining variation, illumination difference or background texture patterns in the smear images. Without expert-annotated ground truth regions, it is not possible to claim that the model is detecting only sickle-shaped cells. The Grad-CAM results are therefore presented as qualitative evidence of possibly attending to clinically relevant morphological patterns, not as definitive proof of interpretability. Limitations and challenges Due to the limited source of datasets, the current dataset employed, though well-annotated and balanced, is from one institution, and this could constrain generalizability to various populations and imaging settings. In addition, although the task at hand is binary classification, real-world applications tend to require multi-class classification of other red blood cell abnormalities. Working on class imbalances is required addition efforts and timing. Another important limitation is the difficulty in acquiring large-scale clinical datasets with reliable annotations. Conclusion and future works The work has proposed an Intuitionistic Fuzzy driven Retinex (IFDR), which is a mathematical model for microscopic image enhancement on SCD. It can adjust the contrast, sharpness, light, and illumination for each image locally. The model outputs clear cell boundaries and distinct shapes, enabling effective differentiation between normal and sickle cells. Followed by this, a DenseNet121-based model is utilized for sickle cell disease classification, effectively learning the differences between normal and sickle cells using the enhanced images. These enhanced images lead the way for deep learning models to learn well with the sickle cell features map and produce increased classification accuracy for sickle cell identification compared to the image without enhancement, as well as conventional methods like Intuitionistic Fuzzy and SSR. The proposed results were compared with various pre-trained models and state-of-the-art methods, demonstrating that the current approach outperforms existing techniques across most of the evaluation metrics. The future work focuses on the abnormal RBC cell automatic counting and segmentation, which will give a quantitative measure. Additionally, the ratio of abnormal to normal cells provides an objective marker of disease severity, symptoms, survival rate, and progression. We plan to collaborate with hematopathologists in future studies to obtain expert annotations and conduct qualitative as well as quantitative validation of the Grad-CAM attention regions. Author contributions All authors have contributed significantly to the conception design and implementation. The data collection and literature survey is done by the Jayasurya and Sriramakrishnan Pathmanaban. Jayasurya and Vinod Kumar did the design and development. Sriramakrishnan Pathmanaban and Jayasurya have contributed to the evaluation and analysis of the model result. Jayasurya and Sriramakrishnan Pathmanaban wrote the manuscript. Each author has contributed sufficiently to the work to accept public responsibility for appropriate portions of the content. The final manuscript was read and approved by all authors. Funding Open access funding provided by Amrita Vishwa Vidyapeetham. Data availability All data generated or analysed during this study are included in this published article. The dataset was collected from Kaggle website ( [https://www.kaggle.com/datasets/florencetushabe/sickle-cell-disease-dataset](https:/www.kaggle.com/datasets/florencetushabe/sickle-cell-disease-dataset) ). Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement All data generated or analysed during this study are included in this published article. The dataset was collected from Kaggle website ( [https://www.kaggle.com/datasets/florencetushabe/sickle-cell-disease-dataset](https:/www.kaggle.com/datasets/florencetushabe/sickle-cell-disease-dataset) ).