Enhancing Esophageal Cancer Detection via Virtual NBI: A Novel Spectrum-Aided Vision Enhancer (SAVE) and Deep Learning Framework

Enhancing Esophageal Cancer Detection via Virtual NBI: A Novel Spectrum-Aided Vision Enhancer (SAVE) and Deep Learning Framework | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Enhancing Esophageal Cancer Detection via Virtual NBI: A Novel Spectrum-Aided Vision Enhancer (SAVE) and Deep Learning Framework Yu-You Tsai, Kun-Hua Lee, Arvind Mukundan, Riya Karmakar, Yaswanth Nagisetti, and 4 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8677211/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 10 You are reading this latest preprint version Abstract Esophageal cancer is a highly aggressive malignancy where early detection is critical for survival. However, early-stage lesions typically present subtle mucosal changes that are difficult to identify using standard White Light Imaging (WLI), and hardware-based Narrow Band Imaging (NBI) is not universally available. In this study, we propose a novel image processing algorithm termed the Spectrum-Aided Vision Enhancer (SAVE) to address these challenges in computer-aided diagnosis (CAD). Leveraging hyperspectral data principles, SAVE transforms standard WLI endoscopic images into enhanced, NBI-like representations, significantly improving mucosal contrast and lesion visibility without requiring additional hardware. To validate the efficacy of this approach for medical image analysis, we utilized a dataset of Squamous Cell Carcinoma (SCC) and dysplasia. We conducted a comprehensive comparative analysis using five state-of-the-art deep learning models: YOLOv8, InceptionV3, Inception-ResNet-V2, ConvNeXt-V2, and MobileNetV2. Experimental results demonstrate that models trained on SAVE-enhanced images significantly outperform those trained on traditional WLI in both classification and detection tasks. This study presents a cost-effective, software-driven solution that integrates advanced image processing with deep learning, offering a robust tool for the automated screening of esophageal malignancies. Biological sciences/Cancer Biological sciences/Computational biology and bioinformatics Physical sciences/Engineering Physical sciences/Mathematics and computing Deep Learning Medical Image Processing Esophageal Cancer Spectrum-Aided Vision Enhancer (SAVE) Virtual NBI Computer-Aided Diagnosis (CAD) Object Detection Figures Figure 1 Figure 2 Introduction Gastroenterology is a branch of medicine that studies the digestive system, its disorders, treatment, diagnosis, and management of gastrointestinal bleeding. Gastrointestinal (GI) tract cancer encompasses a variety of cancers which can cause GI bleeding and are life-threatening. Some of them are stomach cancer, esophageal cancer, colon cancer, small intestine cancer, and others [1]. According to recent statistics, esophageal cancer is the seventh most common cancer, with 572,000 new cases annually, and it ranks 6th in mortality with more than 509,000 deaths. The highest regional reports are indicated in Eastern Asia, with Mongolia and China mentioned among the top five worldwide [2]. Squamous cell carcinoma (SCC) and esophageal adenocarcinoma (EAC) are the two main histological subtypes of this male-dominant aggressive malignancy, each indicates different geographical and racial distribution patterns [3]. SCC is the most common subtype, which is mostly reported in the region of Asia and Africa, whereas EAC is more frequently observed in northern America and Europe and is associated with obesity and gastroesophageal reflux disease (GERD) [4]. Epidemiological research has shown that smoking and drinking alcohol are a significant risk factor for SCC. The two main abnormalities of esophageal cancer, which are dysplasia, and SCC have their own character and identification. SCC mostly affects the middle third of the esophagus and less frequently it can also affect the lower part. However, the cases are rare in the upper esophagus. Squamous dysplasia is a histological lesion kept within the epithelium, characterized by both architectural and cytologic abnormal growth. Architectural changes are described as the loss of cell polarity and lack of surface maturation. Cytologic abnormalities include enlarged nuclei, hyperchromasia (black stained nuclei), pleomorphism (variation in cell shape and size), and increased or abnormal mitotic activity [5]. Metaplasia is a process in which the normal squamous cell layer of the esophagus is changed to columnar cells. Over the past decade, artificial intelligence (AI) applications especially machine learning (ML) have achieved excellent advancement in the field of computer vision [6]. The study focused on classifying the early stages of esophageal cancer is utilizing ResNet50 and a support vector machine (SVM) to classify Barrett’s esophagus (BE) and esophagitis [7]. In the other study done in 2021, the Inception-v2 model was used to classify GI computed tomography (CT) images to diagnose esophageal cancer [8]. To detect and classify various types of esophageal tissue abnormalities white light imaging has been a very common endoscopic technique used for medical image analysis and showed less accurate results than computer-aided detections CADe [9]. In the comparison study of the diagnostic capacity of convolutional neural networks (CNN) and endoscopists, the result demonstrates that CNN is showing better results in detecting early cancer. CNN was organized by utilizing 13,584 endoscopic images from 2639 lesions of cancer and the result is curated using a test dataset of 2940 images from 140 classes [10]. In the performance comparison of AI versus expert endoscopy, images of 23,746 from 1544 cases of esophageal SCCs and images of 4587 from 458 noncancerous tissues and normal tissues were utilized to build an AI system. Since, the accuracies, specificities, and sensitivities for the identification of SCC were high in percentage, the AI system performed well [11]. The major limitations of WLI shown in these comparisons include being less sensitive to point out the mucosal and vascular changes that are indicative of early or serious stages of cancer, using broad-spectrum light which doesn’t use specific light wavelength to identify the lesions clearly, and having a longer learning curve [12]. However, most of the studies use WLI images which utilize three primary color bands, integrating these highly performed AI systems with hyperspectral imaging (HSI) is an innovative way of capturing images beyond WLI and brings high assurance in medical fields. For instance, the five-layered CNN with the dataset of 300 HSI was trained and fine-tuned then it achieved 94.3% accuracy [13,14]. HSI which captures and processes information across the electromagnetic spectrum is an emerging technique for medical imaging [15]. Basically, HSI operates by capturing images based on wavelength variation using spectroscopic techniques. The photons scattering in the blood vessel is recorded to produce spectral images across narrow spectral bands, particularly resulting in multiple bands. For specific bands, the relative amount of light absorbance or reflectance which helps to increase the contrast of cancerous tissues for each image will be done. This information is compiled into a finite 3D volume with two spatial dimensions (X rows and Y columns) and one spectral dimension (λ wavelength), known as a hyperspectral image, hyperspectral (HS) cube, or hypercube [16,17]. The working principle of an HSI that makes use of particular cameras for implementation purposes is divided into main 3 techniques: spectral scanning, spatial scanning, and snapshot method [18]. However, these three techniques are expensive and not easy to use because they require highly specialized cameras such as pushbroom which are more complex than normal imaging systems. They also need high-quality optics for the spectrometer or filters to separate and capture specific wavelengths, and they use precise calibration to ensure accurate data. Hence, the more specific technique called narrow-band imaging (NBI) which is obtained from HSI conversion, helps in getting more convenient and high-resolution images. NBI is an endoscopic method that makes use of green and blue lights to highlight mucosal and submucosal vasculature [19]. The shorter wavelength of these visible lights (415 nm and 540 nm ) has low tissue permeability which makes it a highly preferable way for identifying mucosal surface structure [20,21]. Because of the strong ability to be absorbed by hemoglobin and its low penetration behavior blue light is used effectively in superficial layers of mucosa. The green light has a slightly longer wavelength and can penetrate a bit deeper than the blue light utilized to visualize the submucosal blood vessel [22]. This NBI method was described in 2003 for the first time and helps to facilitate the detection of early SCC [23]. The study which was done in 1780 esophageal cancer images that include 935 NBI images and 845 WLI, NBI showed better results for the SCC group and good estimation for the normal group in comparison [24]. Similar research was done at the Kaohsiung Medical University for 45 patients and proved that the diagnosis of NBI has greater sensitivity compared with that of white light endoscopy [25]. Therefore, in this work, NBI has been combined with HSI and used to select a few spectrums to enhance the contrast of esophageal cancer identification, particularly dysplasia before it turns into SCC. The CAD system is the best way to increase the visualization of esophageal tissue and enable accurate differentiation among normal, dysplasia, and SCC especially if it integrates with SAVE. By applying the SAVE model and training different ML models for the given dataset, this work focuses on creating a robust model that can identify and prevent pre-cancerous conditions. For the given dataset the captured WLI images are converted into SAVE and made available for further training. Methods 2.1 Dataset The dataset that is used in this study is sourced from Kaohsiung Armed Forces General Hospital, comprising a total of 5370 images. The given 5370 images are captured using an Olympus endoscope (CV-290, Olympus), and the WLI images are converted into SAVE images to get high performance of the ML models and identify early-stage esophageal cancer. The dataset is comprised of 2 main types of esophageal abnormalities and the normal class for better results. Dysplasia is the earliest stage of esophageal cancer and SCC refers to the serious stage of the tumor. For the model training purpose, the given dataset was split into a train, validation, and test with a total number of images divided into training, validation, and testing sets with a ratio of 70:20:10, respectively. Each split of the dataset contains three main classes and specifically the training dataset contains 3 classes with image distribution of 1420, 1175, and 1160 for the dysplasia, normal, and SCC respectively for both WLI and SAVE images (overall distribution of images per class is summarized in Table S2). All methods were carried out in accordance with relevant guidelines and regulations. The experimental protocols involving human subjects were approved by the Institutional Review Board of Kaohsiung Armed Forces General Hospital (Approval No. KAFGHIRB 114 − 022). Due to the retrospective nature of the study and the use of de-identified images, the requirement for informed consent was waived by the Institutional Review Board of Kaohsiung Armed Forces General Hospital. 2.2 SAVE This study introduces the Spectrum-Aided Vision Enhancer (SAVE), a novel image processing technique designed to transform standard White Light Imaging (WLI) into hyperspectral-enhanced representations (HSI and SAVE images). The primary objective of SAVE is to augment the diagnostic capability of conventional endoscopy by recovering detailed spectral information that is typically lost in traditional RGB imaging. The core of the SAVE mechanism lies in the precise calibration of endoscopic RGB data to match the reflectance spectra obtained from a spectrometer. To establish a robust mapping between the WLI images and the ground-truth spectral data, we utilized a Macbeth Color Checker (X-Rite Classic), which contains 24 standardized color patches representing natural reflectance spectra. The calibration pipeline begins by converting the raw endoscopic images from the standard sRGB color space into the CIE 1931 XYZ color space, a device-independent standard in colorimetry where X, Y, and Z correspond to the trichromatic response functions. To ensure high-fidelity color reproduction, we applied a series of correction algorithms to address common imaging artifacts, including non-linear sensor response, dark current noise, imperfect color separation, and lens distortion. The corrected XYZ values (XYZ correct ) are derived through this rigorous calibration process. The transformation from raw RGB to the initial XYZ space is formulated in Eq. ( 1 ), while the subsequent error correction and calculation of the final XYZ correct are defined in Eq. ( 2 ). $$\:\left[C\right]=\left[XY{Z}_{Spectrum}\right]\times\:pinv\left(\left[V\right]\right)$$ 1 $$\:\left[XY{Z}_{Correct}\right]=\left[C\right]\times\:\left[V\right]$$ 2 Following the initial color space conversion, the calibration process necessitates a rigorous alignment between the generated SAVE representations and the clinical gold standard—Narrow Band Imaging (NBI) captured by Olympus endoscopes. To achieve this, we employed Principal Component Analysis (PCA) to extract the most significant features from the high-dimensional reflectance spectra. This step is crucial for reducing data dimensionality while preserving essential spectral information, thereby optimizing computational efficiency. Our PCA analysis revealed that the first six principal components were sufficient to capture 99.64% of the total variance in the spectral data. This finding confirms that a low-dimensional representation can effectively model the complex reflectance properties of esophageal tissue. To validate the accuracy of our color calibration, we calculated the Root Mean Square Error (RMSE) between the corrected image values (XYZ correct ) and the ground-truth spectrometer measurements (XYZ spectrum ). The resulting mean RMSE of 0.19 indicates a minimal color difference, demonstrating the high fidelity of our calibration pipeline. The successful calibration serves as the foundation for converting standard WLI images into Hyperspectral Imaging (HSI) formats and subsequently into virtual NBI (SAVE) images. To formalize this conversion, we conducted a multiple regression analysis to model the relationship between the input WLI data and the target spectral components. This analysis yielded a transformation matrix, defined in Eq. ( 3 ), which serves as the mathematical kernel for mapping the broad-spectrum WLI signals to the specific narrow-band wavelengths required for SAVE. By applying this matrix, we can synthetically reconstruct tissue visualization that mimics the contrast-enhancement properties of hardware-based NBI, effectively highlighting vascular structures and mucosal patterns. $$\:\left[M\right]=\left[Score\right]\times\:pinv\left(\left[{V}_{Color}\right]\right)$$ 3 The analog spectrum derived from the corrected XYZ values Equation. 4 demonstrated minimal color discrepancies compared to real measurements, supporting the system’s accuracy: $$\:{\left[SSpectrum\right]}_{380\sim780nm}=\left[EV\right]\left[M\right]\left[{V}_{Color}\right]$$ 4 Upon reconstructing the hyperspectral data from RGB inputs, the subsequent phase involved synthesizing virtual NBI representations and validating their fidelity against clinical standards. We utilized the Olympus NBI system as the ground truth benchmark, comparing its output directly with our algorithmically generated SAVE images. This comparison was facilitated using the 24-patch color checker to ensure consistent colorimetric evaluation. To quantify the perceptual similarity between the synthetic SAVE images and the hardware-captured NBI images, we computed the CIEDE2000 color difference metric (ΔE 00 ) for each color patch. The initial analysis yielded a mean ΔE 00 of 2.79, indicating a high degree of visual resemblance with only minor deviations. These discrepancies were primarily attributed to inherent differences in the illumination spectra, reflectance properties, and color-matching functions between the simulated environment and the physical endoscopic hardware. A critical observation was that the spectral mismatch was most pronounced in the 450–540 nm wavelength range, which is vital for visualizing superficial mucosal capillaries. To mitigate this and further minimize the color divergence, we implemented a secondary calibration layer utilizing the Cauchy-Lorentz distribution. This statistical model, defined in Eq. ( 5 ), is particularly effective for characterizing spectral resonance and was applied to fine-tune the system's spectral response. By adjusting the distribution parameters, we significantly reduced the spectral deviation, ensuring that the SAVE images provided a clinically accurate emulation of the NBI modality. $$\:f\left(x;{x}_{0},\gamma\:\right)=\frac{1}{\pi\:\gamma\:\left[1+{\frac{\left(x-{x}_{0}\right)}{\gamma\:}}^{2}\right]}=\frac{1}{\pi\:}\left[\frac{\gamma\:}{\left(x-{x}_{0}\right)2+{\gamma\:}^{2}}\right]$$ 5 To refine the light spectrum parameters, we employed the dual annealing optimization algorithm. This stochastic approach integrates the robustness of Classical Simulated Annealing (CSA) with the efficiency of Fast Simulated Annealing (FSA), allowing for a global search of the optimal spectral configuration. Following this optimization, the mean CIEDE2000 color difference was stabilized at 5.36. While this value is slightly higher than the initial calibration, it accounts for the complex post-processing inherent in the Olympus system—specifically, the introduction of brownish tones alongside the primary hemoglobin absorption peaks (415 nm and 540 nm) to enhance tissue realism. To rigorously assess the fidelity of the SAVE-generated images against the gold-standard NBI, we utilized a triad of image quality metrics: Structural Similarity Index Measure (SSIM), Entropy, and Peak Signal-to-Noise Ratio (PSNR). The SSIM analysis yielded a similarity score of 94.27%. This high correlation confirms that the SAVE mechanism successfully preserves the structural details and textural information of the mucosal surface, making it a reliable surrogate for hardware NBI. Entropy was calculated to evaluate the richness of image information. The results showed a marginal difference of only 0.37% between the SAVE and Olympus NBI images, indicating that our method maintains the textural complexity required for diagnosis. The PSNR, a standard metric for reconstruction quality, reached 27.88 dB. This value suggests that the SAVE images retain high signal fidelity with minimal noise introduction during the transformation process. The comprehensive workflow, from initial RGB input to the final optimized SAVE output, is schematically illustrated in Fig. 1 . 2.3 DL algorithm 2.3.1 YOLOv8 The first YOLO model was introduced in 2015 by Joseph Redmon in a C-based repository named Darknet during his Ph.D. studies at the University of Washington. Since then, the community has continued developing and enhancing the subsequent YOLO version [26,27]. Ultralytics the community that introduced YOLOv5 released YOLOv8 in January 2023. This new version supports different computer vision tasks including object detection, object segmentation, tracking, and classification. By targeting the specific need of the task YOLOv8 developed with five different variants including YOLOv8n (nano), YOLOv8s (small), YOLOv8m (medium), YOLOv8l (large), and YOLOv8x (extra-large) [28,29]. The backbone responsible for the feature extraction of the input images is one of the main components of the YOLOv8 architecture. By making use of the path aggregation network (PANet) with feature pyramid network (FPN) the other architecture called neck detects objects at various levels(an illustration of the YOLOv8 architecture used in this study is shown in Figure S1 ). The detection head of YOLOv8 is where the final prediction is done and for tasks like object classification and regression tasks decoupled head is utilized [30,31]. For different tasks, YOLOv8 uses different Loss functions. The classification loss utilizes Variational Focal Loss (VFL), and the regression employs a combination of Distribution Focal Loss (DFL) and Complete IoU (CIoU) Loss. Unlike Focal Loss (FL) and Quality Focal Loss (QFL), which are symmetric, VFL uses an asymmetric weighting technique to find out the solution of the imbalance between negative and positive samples. $$\:VF{L}_{\left(p,q\right)}=\left\{-q\left(q\:\text{log}\text{log}\left(P\right)+\left(1-q\right)\text{log}\left(1-p\right)\right),q>0-\alpha\:{P}^{y}\text{log}\left(1-p\right),q=0\right\}$$ 6 Where VFL stands for variational focal loss, p represents the label and q is a value calculated by normalized alignment matric if a positive sample is taken. 2.3.2 InceptionV3 Transfer learning is a way of solving a new problem based on the previous learning history of the model. The main characteristic of the neural network is that the layer that is closer to the output is updated based on the new data, while the remaining hidden layers are kept unchanged [32]. InceptionV3 is one of the transfer learning models built as an advanced version of Google Net (InceptionV1) which achieved the first prize in the 2014 ILSVRC Competition [33]. This model is the other version of Inception-V1 and Inception-V2 architectures and it’s deep CNN which is trained on low-configuration computers [34]. Moreover, the model is trained on the large-scale ImageNet dataset and has the ability to classify images into 1000 different groups [35]. The architecture of inceptionV3 can be divided into smaller convolutional kernels which has a huge effect of lowering the model’s parameter and minimizing the probability of overfitting [36]. The main architecture includes convolution, MaxPool, Average Pool, fully connected, dropout, and SoftMax [37]. The loss function of the inceptionV3 for the classification task is represented in Eq. 7 . $$\:L=-\:{\sum\:}_{i=1}^{c}{y}_{i}\text{log}\left({x}_{i}\right)$$ 7 Where y i is the true value of class i, x i is the predicted probability for class I, and C is the overall number of classes. 2.3.3 ConvNeXt-V2 ConvNeXt-V2 is a modern convolutional neural network model that aims at closing the performance gap between traditional CNN models and models based on transformers. It uses the concepts of the self-attention mechanism of transformers by a Fully Convolutional Masked Autoencoder (FCMAE) framework, which allows the model to learn more expressive and strong image features. Under this framework, portions of the input image are masked, and the network learns to replicate the missing portions, which enhances its feature-learning capacity. To facilitate this process in ConvNeXt-V2, regular convolutions in the masked regions are substituted with sparse convolutions, and the reconstruction loss of the pre-training is based on Mean Squared Error (MSE). Another major advancement in ConvNeXt-V2 is the integration of Global Response Normalization (GRN), a mechanism designed to mitigate feature collapse—a phenomenon where deep models saturate or deactivate feature maps, limiting their representational capacity. GRN enhances channel-wise feature competition by normalizing the global activation strength, thereby stabilizing the learning process. Mathematically, for a given input feature map with defined height, width, and channel dimensions, the global response is first computed by aggregating the spatial information across each channel using the L2-norm, as defined in Eq. ( 8 ) [38]: $$\:G\left(X\right)=\left\{\left|\left|{X}_{1}\right|\right|,\:\left|\left|{X}_{2}\right|\right|,\dots\:,\left|\left|{X}_{c}\right|\right|,\right\}$$ 8 Subsequently, a channel-wise normalization step is applied to compute the relative importance of each channel, denoted as N(X), as shown in Eq. ( 9 ): $$\:N\left(\left|\left|{X}_{i}\right|\right|\right)=\frac{\left|\left|{X}_{i}\right|\right|}{\sum\:_{j=1}^{C}\left|\left|{X}_{j}\right|\right|}$$ 9 Finally, the calibrated feature map is obtained by scaling the original input with learnable scaling and shifting parameters, ensuring that the network retains the flexibility to adaptively emphasize informative features. This transformation is expressed in Eq. ( 10 ): $$\:{X}_{i}^{{\prime\:}}=\gamma\:\bullet\:{X}_{i}\bullet\:N\left(G\left({X}_{i}\right)\right)+\beta\:+{X}_{i}$$ 10 Where \(\:\gamma\:\) and \(\:\beta\:\) are learnable parameters ConvNeXt-V2 was employed in this work to classify images of the esophagus (Normal, Dysplasia, and SCC) with the help of WLI and SAVE-generated hyperspectral band-selected data. Resizing was done using all the images to the size of 224x 224 pixels prior to training. Due to the design of ConvNeXt-V2, focusing on high-accuracy classification with high generalization, the algorithm is compatible with medical images where the slightest changes in tissue texture and color can need to be detected. In order to guarantee the dependability of the assessment, 5-fold cross-validation was used, and the average performance scores across all folds were obtained. This enabled the model to be tested equally on WLI and the hyperspectral inputs. 2.3.4 Inception-ResNet-V2 Inception-ResNet-V2 is a deep convolutional neural network, which combines the multi-scale feature extracting capability of Inception modules with the stability of residual learning. The model starts with a stem block which is capable of basic feature extraction and then there are three major blocks, Inception-ResNet-A, Inception-ResNet-B, and Inception-ResNet-C. Many convolutional branches operating in parallel are contained in each block, and this aids the network to acquire features of images at various scales. A defining characteristic of this architecture is the incorporation of shortcut connections (skip connections) within each Inception block. These connections compel the network to learn a residual mapping rather than a direct transformation, which significantly enhances training stability and mitigates the vanishing gradient problem in deep networks. The residual learning mechanism is formally defined in Eq. ( 11 ) [39]: $$\:H\left(x\right)=F\left(x\right)+x$$ 11 Where 𝑥 represents the input to the block, whereas 𝐹(𝑥) is the multi-branch convolutional output. All the blocks in the architecture are done with batch normalization and reduction in order to accelerate training and regulate model complexity. The Inception-ResNet-V2 has a number of hundreds of layers, which enables it to make both fine and complex contributions to medical images. The model was applied in this research to categorize esophageal images of the WLI and SAVE. All the images were downscaled to 224 x 224 pixels and then submitted to the network. To make the evaluation reliable, a 5-fold cross-validation procedure was performed, i.e. the dataset was split into five equal parts and model was trained and validated five times. This increased the performance comparison between the WLI and the hyperspectral images in terms of being more accurate and unbiased. 2.3.5 MobileNetV2 MobileNetV2 is one of the neural networks which is developed by Google and it runs on any device efficiently with less computational power and gives high accuracy [40]. The architecture of the model contains an inverted residual block, linear bottleneck layer, depth-wise separable convolution, ReLU6 Activation, expansion layer (1X1 convolution), stridden convolutions for down-sampling, global average pooling (GAP), and final fully connected layer [41,42]. The model depends on depth-wise separable convolutions which separate standard convolution into two steps: depth-wise (pre-channel filtering) and pointwise (combining information across channels). This is important to reduce the computational complexity compared to traditional convolutions while maintaining accuracy [43]. The other architecture of mobilenetv2 which are inverted residual and linear bottleneck structures significantly reduce the computational cost of convolutions making it both efficient and effective to use [44]. Results Table 1 Performance of various ML models for a given dataset. Models Imaging Modalities Classes Precision Recall F1 Score Specificity YOLOv8n WLI Normal 78 87 82 85 Dysplasia 93 85 89 97 SCC 87 82 84 97 SAVE Normal 78 87 82 85 Dysplasia 90 88 89 96 SCC 86 76 81 94 InceptionV3 WLI Normal 84 93 88 89 Dysplasia 91 88 90 96 SCC 97 88 92 98 SAVE Normal 89 95 92 92 Dysplasia 92 97 94 96 SCC 100 85 92 100 Inception-ResNet-V2 WLI Normal 90 93 92 94 Dysplasia 89 94 91 94 SCC 94 85 89 97 SAVE Normal 97 90 94 98 Dysplasia 86 94 90 93 SCC 91 91 91 96 ConvNeXt-V2 WLI Normal 78 88 83 85 Dysplasia 96 74 83 98 SCC 89 97 93 94 SAVE Normal 93 86 90 96 Dysplasia 89 99 94 94 SCC 92 91 92 96 MobileNetV2 WLI Normal 85 95 90 89 Dysplasia 88 88 88 88 SCC 97 82 89 89 SAVE Normal 83 98 90 88 Dysplasia 91 88 90 96 SCC 100 82 90 100 The cross-validation that was performed in five-fold in all five deep learning models, i.e., YOLOv8n, InceptionV3, Inception-ResNet-V2, ConvNeXt-V2, and MobileNetV2, demonstrated different trends in the diagnostic performance of the models with application to WLI and SAVE imaging modalities. This variation in precision, recall, F1-score, and specificity between the Normal, Dysplasia as well as SCC classes can be directly explained by the variation in model architecture and the quality of visual information that each imaging modality possesses. SAVE scans always provided better clarity, contrast, and structural visibility, which contributed greatly to the change of the pattern of performance in terms of classes. Fold-wise numerical results are given in Supplementary Tables S1-S25 and all fold-wise loss plots and confusion matrices are found in Figures S1 -S90. The first one began with YOLOv8n whose lightweight structure was relatively sensitive to the gains made by SAVE imaging. There was a minor improvement in dysplasia performance under SAVE mainly because of the increase in the level of illumination and clarity of the epithelial pattern by which early-stage morphological variations, which are weak in WLI, can be detected. This has led to more increase in recall demonstrating that a smaller number of dysplastic cases were overlooked. But SCC recall declined with SAVE as greater strengthening of intermediate dysplastic features by the modality introduced minor overlaps between high-grade dysplasia and early SCC leading to misclassification at the border between the two classes. Normal class was also similar in modalities as YOLOv8n already works well with normal tissue and the extra SAVE clarity did not dramatically change the property of the class. These results indicate the consistency of YOLOv8n and also exemplify its weaknesses in discriminating closely related pathological classes in the case of adding visual complexity through SAVE. Fold-wise numerical results are given in Supplementary Tables S11-S15 and all fold-wise loss plots and confusion matrices are found in Figures S71-S90. InceptionV3, on the contrary, was more advantaged by SAVE images. Its multi-scale convolutional filters are based heavily on rich spatial detail and color gradient both of which are more exhibited in SAVE. The recall of dysplasia (strong) and near-perfect recall showed an improvement of values under SAVE, which highlights the fact that the latter is more accurate in the detection of small lesions. Equally, precision in SCC was 100% with SAVE, showing that more evident morphological indicators resulted in the model being able to remove false positives completely. The sharpening of the epithelial edges, and the even distribution of light in SAVE also led to improved normal tissue recognition since the model could avoid misclassifying benign abnormalities that were a common occurrence in WLI. Collectively, these improvements affirm that the InceptionV3 architecture is extremely compatible with the sophisticated spectral and structural data delivered by SAVE which allows to perform more confident classification in all tissue types. Fold-wise numerical results are given in Supplementary Tables S1-S5 and all fold-wise loss plots and confusion matrices are in Figures S21-S40. Inception-ResNet-V2 also produced good and robust performance in both modalities, however, SAVE provided some additional benefits that enhanced the reliability of the classification. This architecture has residual connections together with inception modules which makes it especially effective at refining intermediate features. The highest improvement was observed with the Normal class with SAVE whereby the precision was enhanced significantly. This was the case because SAVE lowers the appearance of shadows, glare and minor inflammatory alterations which may appear like pathology and enhances the confidence of the model to recognize non-pathological mucosa. Dysplasia performance was also very strong but the precision was very slight because SAVE tended to exaggerate benign textural changes which made the model over-detect dysplasia in normal areas. It was observed that SCC performance was also balanced, with both precision and recall improving or retaining high levels due to the fact that SCC morphology is refined obviously in SAVE and with ease in the deep layers of the feature extraction. The findings highlight the versatility of the Inception-ResNet-V2 that successfully used the better visual cues of SAVE to sharpen the predictions without causing instability. Fold-wise numerical results are given in Supplementary Tables S16-S20 and all fold-wise loss plots and confusion matrices are in Figures S61-S70. ConvNeXt-V2 model showed the most drastic improvements with SAVE and particularly with dysplasia class. This model is based on the transformer-style architectural ideas and relies on the detailed texture patterns and long-range dependencies which are better offered by SAVE imaging than by WLI. Recall of dysplasia increased significantly when under WLI, it was low and when under SAVE, it was almost perfect, which is that ConvNeXt-V2 needs high-quality structural and textural detail in order to identify early or mild lesions. Normal and SCC classes were also benefitted and displayed more equalized values of precision and recall with SAVE. WLI weaknesses of irregular brightness, contrast, and lesser expression of gentle architectural distortion were barriers to the generalizability of this architecture in some classes. SAVE addressed these errors and enabled ConvNeXt-V2 to see the tissue patterns in a more holistic and correct way. These findings prove that highly vision transformer-like models are best used in cases of improved imaging conditions. Fold-wise numerical results are given in Supplementary Tables S21-S25 and all fold-wise loss plots and confusion matrices are in Figures S1 -S20. Lastly, even though MobileNetV2 is lightweight, it also showed significant gains with SAVE. This model suffers a shortcoming of not being able to extract features of WLI limited features of distinguishing dysplasia and borderline pathology. To counter this, SAVE made available more explicit micro-structural characteristics that made the boundaries of decisions easier to find in the model. The Dysplasia F1-scores were better and SCC precision was 100% indicating that MobileNetV2 was able to reliably discriminate between advanced malignant appearances when assisted by high-quality imaging. There was also the improvement of the Normal class as SAVE reduced noise, asymmetric lighting, and other minor mucosal artifacts that previously led to a false positive. Therefore, MobileNetV2 has shown that also small models can significantly benefit the added structural and spectral cues of SAVE that can be used to achieve performance in some classes similar to much larger designs. Fold-wise numerical results are given in Supplementary Tables S6-S10 and all fold-wise loss plots and confusion matrices are found in Figures S41-S60. Discussion The result of this study demonstrates that the SAVE model has better results than the conventional one. Hence by enhancing the early identification of esophageal cancer, the SAVE model can contribute to better patient outcomes, as identifying and treating early cancer can prevent the spread of cancer into the severe stage. Kaohsiung Medical University Chung-Ho Memorial Hospital used as the only source of the dataset which could limit the model’s generalization. Incorporating data from different sources like multiple hospitals, different regions, races, and age groups would help to consider the geographical and genetic variability of esophageal cancer and races could increase the model's capacity to find esophageal cancer across various demographic groups. Even though the scope of the dataset is somewhat limited to identifying and diagnosing all the types of esophageal cancer, this study aimed to establish a benchmark for ML models applied to esophageal cancer detection. Having research done on SCC which is the main type of esophageal cancer type is very crucial, as it represents both ends of the spectrum early, hard-to-detect cases (dysplasia), and severe, late-stage cases of the disease (SCC). the other important types of esophageal cancer that are not mentioned in this work include Barrett's esophagus and high- and -low-grade dysplasia. Barrett's esophagus is a stage in which the normal lining of the esophagus is replaced with the other abnormal tissue and this is the result of long-term acid reflux most of the time. the word low-grade dysplasia refers to starting early abnormal growth of the esophagus lining but it doesn't directly threaten human life and has a chance of being treated and cured. the highest stage of low-grade dysplasia is called high-grade dysplasia. the inclusion of these different varieties of cancer in the study and testing the performance of the SAVE model toward the WLI is the future direction of this research. For comparing the performance of the SAVE model with the WLI model and analyzing different architecture's performance in the identification of different types of esophageal cancer, five different ML models were utilized. this comparative technique makes the current study a great benchmark for future studies that may include many more ML models and a larger amount of dataset. the identification of which ML model is more effective in medical image tasking and which model needs further fine-tuning or optimization is done by current analysis. the name of the models which was trained for a given dataset includes, YOLOv8, InceptionV3, ConvNeXt-V2, Inception-ResNet-V2, and MobileNetV2. Out of all the tested models, YOLOv8 showed the best result in comparison with the other four models on the basis of accuracy and reliability. Moreover, this study focuses on sensitivity (recall), precision, and f1-score, future work will incorporate ROC curves analysis to better examine the balance between sensitivity and specificity and to provide AUC-based performance metrics. By addressing this the applicability and reliability of the SAVE model could future improve in the medical sector and lead to accurate identification of esophageal cancer. Conclusions In this study, we proposed and validated a novel image enhancement algorithm, the Spectrum-Aided Vision Enhancer (SAVE), designed to overcome the limitations of standard White Light Imaging (WLI) in esophageal cancer detection. By leveraging hyperspectral principles to generate virtual Narrow Band Imaging (NBI) representations, the SAVE mechanism significantly improves the visibility of mucosal vasculature and subtle lesions without the need for specialized hardware. This software-driven approach provides a cost-effective solution for enhancing image quality, which is a critical preprocessing step for automated medical diagnosis. Our comprehensive evaluation using five state-of-the-art deep learning models—YOLOv8, InceptionV3, Inception-ResNet-V2, ConvNeXt-V2, and MobileNetV2—demonstrates that the integration of SAVE consistently outperforms traditional WLI-based training. The models trained on SAVE-enhanced datasets exhibited superior accuracy, sensitivity, and specificity in classifying and detecting Squamous Cell Carcinoma (SCC) and dysplasia. These findings confirm that domain-specific image enhancement can effectively augment the feature extraction capabilities of deep neural networks, leading to more robust and reliable Computer-Aided Diagnosis (CAD) systems. Future work will focus on expanding the dataset to include a wider variety of esophageal pathologies and conducting multi-center clinical trials to further validate the generalizability of the SAVE algorithm. Additionally, we aim to optimize the computational efficiency of the SAVE mechanism for real-time deployment in clinical endoscopic systems. Ultimately, this study highlights the potential of combining advanced image processing techniques with deep learning to assist endoscopists in early cancer detection, potentially reducing missed diagnoses and improving patient outcomes. Declarations Institutional Review Board Statement The study was conducted in accordance with the Declaration of Helsinki and approved by the Institutional Review Board of Kaohsiung Armed Forces General Hospital (Protocol code KAFGHIRB 114 − 022; approved on 25 April 2025). The requirement for informed consent was waived by the Institutional Review Board due to the retrospective design of the study. Conflicts of Interest: The authors declare no conflicts of interest. Funding: This research received support from the National Science and Technology Council, Republic of China through the following grants: NSTC 113-2221-E-194-011-MY3. Additionally, financial support was provided by the Dalin Tzu Chi Hospital, Buddhist Tzu Chi Medical Foundation-National Chung Cheng University Joint Research Program and Kaohsiung Armed Forces General Hospital Research Program KAFGH_D_115-005 in Taiwan. Author Contribution Conceptualization **:** R.K., Y.-Y.T., Y.N, A.M. & H.-C.W. ; data curation : Y.-Y.T., D.G.S, K.-H.L.., R.K., A.M. & H.-C.W. ; formal analysis: D.G.S, K.-H.L., Y.N, A.M., H.-C.W. ; funding acquisition Y.-Y.T., K.-H.L., A.M. & H.-C.W. ; investigation: D.G.S, R.K., K.-H.L.; methodology: D.G.S, R.K., C.-Y.K, & A.M. ; project administration: C.-W.H, A.M. and H.-C.W.; resources, C.-W.H, A.M. & H.-C.W.; software: R.K. and C.-W.H. ; supervision, Y.N, H.-C.W.; writing—original draft, Y.-Y.T., S.L.N. and D.G; writing—review and editing: R.K., S.L.N., A.M. & H.-C.W. All authors have read and agreed to the published version of the manuscript. Data Availability The data presented in this study are available in this article; further considerable requests can be made to the corresponding author (H.-C.W.). Disclaimer/Publisher’s Note : The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. References Fan, S.; Xu, L.; Fan, Y.; Wei, K.; Li, L. Computer-aided detection of small intestinal ulcer and erosion in wireless capsule endoscopy images. Phys Med Biol 2018 , 63 , 165001, doi:10.1088/1361-6560/aad51c. He, Y.; Li, D.; Shan, B.; Liang, D.; Shi, J.; Chen, W.; He, J. Incidence and mortality of esophagus cancer in China, 2008–2012. Chin J Cancer Res 2019 , 31 , 426–434, doi:10.21147/j.issn.1000-9604.2019.03.04. Abbas, G.; Krasna, M. Overview of esophageal cancer. Ann Cardiothorac Surg 2017 , 6 , 131–136, doi:10.21037/acs.2017.03.03. Kojima, T.; Shah, M.A.; Muro, K.; Francois, E.; Adenis, A.; Hsu, C.H.; Doi, T.; Moriwaki, T.; Kim, S.B.; Lee, S.H.; et al. Randomized Phase III KEYNOTE-181 Study of Pembrolizumab Versus Chemotherapy in Advanced Esophageal Cancer. J Clin Oncol 2020 , 38 , 4138–4148, doi:10.1200/JCO.20.01888. Jain, S.; Dhingra, S. Pathology of esophageal cancer and Barrett's esophagus. Ann Cardiothorac Surg 2017 , 6 , 99–109, doi:10.21037/acs.2017.03.06. Barragan-Montero, A.; Javaid, U.; Valdes, G.; Nguyen, D.; Desbordes, P.; Macq, B.; Willems, S.; Vandewinckele, L.; Holmstrom, M.; Lofman, F.; et al. Artificial intelligence and machine learning for medical imaging: A technology review. Phys Med 2021 , 83 , 242–256, doi:10.1016/j.ejmp.2021.04.016. Kumar A, C.; Mubarak, D.M.N. Classification of Early Stages of Esophageal Cancer Using Transfer Learning. Irbm 2022 , 43 , 251–258, doi:10.1016/j.irbm.2021.10.003. Chen, K.B.; Xuan, Y.; Lin, A.J.; Guo, S.H. Esophageal cancer detection based on classification of gastrointestinal CT images using improved Faster RCNN. Comput Methods Programs Biomed 2021 , 207 , 106172, doi:10.1016/j.cmpb.2021.106172. Wang, P.; Liu, P.; Brown, J.R.G.; Berzin, T.M.; Zhou, G.; Lei, S.; Liu, X.; Li, L.; Xiao, X. Lower adenoma miss rate of computer-aided detection-assisted colonoscopy vs routine white-light colonoscopy in a prospective tandem study. Gastroenterology 2020 , 159 , 1252–1261. e1255. Ikenoyama, Y.; Hirasawa, T.; Ishioka, M.; Namikawa, K.; Yoshimizu, S.; Horiuchi, Y.; Ishiyama, A.; Yoshio, T.; Tsuchida, T.; Takeuchi, Y.; et al. Detecting early gastric cancer: Comparison between the diagnostic ability of convolutional neural networks and endoscopists. Dig Endosc 2021 , 33 , 141–150, doi:10.1111/den.13688. Fukuda, H.; Ishihara, R.; Kato, Y.; Matsunaga, T.; Nishida, T.; Yamada, T.; Ogiyama, H.; Horie, M.; Kinoshita, K.; Tada, T. Comparison of performances of artificial intelligence versus expert endoscopists for real-time assisted diagnosis of esophageal squamous cell carcinoma (with video). Gastrointest Endosc 2020 , 92 , 848–855, doi:10.1016/j.gie.2020.05.043. Muto, M.; Minashi, K.; Yano, T.; Saito, Y.; Oda, I.; Nonaka, S.; Omori, T.; Sugiura, H.; Goda, K.; Kaise, M.; et al. Early detection of superficial squamous cell carcinoma in the head and neck region and esophagus by narrow band imaging: a multicenter randomized controlled trial. J Clin Oncol 2010 , 28 , 1566–1572, doi:10.1200/JCO.2009.25.4680. Grigoroiu, A.; Yoon, J.; Bohndiek, S.E. Deep learning applied to hyperspectral endoscopy for online spectral classification. Sci Rep 2020 , 10 , 3947, doi:10.1038/s41598-020-60574-6. Maktabi, M.; Wichmann, Y.; Kohler, H.; Ahle, H.; Lorenz, D.; Bange, M.; Braun, S.; Gockel, I.; Chalopin, C.; Thieme, R. Tumor cell identification and classification in esophageal adenocarcinoma specimens by hyperspectral imaging. Sci Rep 2022 , 12 , 4508, doi:10.1038/s41598-022-07524-6. Khan, M.J.; Khan, H.S.; Yousaf, A.; Khurshid, K.; Abbas, A. Modern Trends in Hyperspectral Image Analysis: A Review. IEEE Access 2018 , 6 , 14118–14129, doi:10.1109/access.2018.2812999. Collins, T.; Maktabi, M.; Barberio, M.; Bencteux, V.; Jansen-Winkeln, B.; Chalopin, C.; Marescaux, J.; Hostettler, A.; Diana, M.; Gockel, I. Automatic Recognition of Colon and Esophagogastric Cancer with Machine Learning and Hyperspectral Imaging. Diagnostics (Basel) 2021 , 11 , doi:10.3390/diagnostics11101810. Ma, J.; Sun, D.W.; Pu, H.; Cheng, J.H.; Wei, Q. Advanced Techniques for Hyperspectral Imaging in the Food Industry: Principles and Recent Applications. Annu Rev Food Sci Technol 2019 , 10 , 197–220, doi:10.1146/annurev-food-032818-121155. Yoon, J. Hyperspectral Imaging for Clinical Applications. BioChip Journal 2022 , 16 , 1–12, doi:10.1007/s13206-021-00041-0. van Schaik, J.E.; Halmos, G.B.; Witjes, M.J.H.; Plaat, B.E.C. An overview of the current clinical status of optical imaging in head and neck cancer with a focus on Narrow Band imaging and fluorescence optical imaging. Oral Oncol 2021 , 121 , 105504, doi:10.1016/j.oraloncology.2021.105504. Kurumi, H.; Nonaka, K.; Ikebuchi, Y.; Yoshida, A.; Kawaguchi, K.; Yashima, K.; Isomoto, H. Fundamentals, Diagnostic Capabilities and Perspective of Narrow Band Imaging for Early Gastric Cancer. J Clin Med 2021 , 10 , doi:10.3390/jcm10132918. Yang, K.-Y.; Mukundan, A.; Tsao, Y.-M.; Shi, X.-H.; Huang, C.-W.; Wang, H.-C. Evaluating Hyperspectral Techniques Using Objective Metrics Research on Analog Narrowband Image. 2023 . Yao, K. Principles of magnifying endoscopy with narrow-band imaging. In Zoom Gastroscopy: Magnifying Endoscopy in the Stomach ; Springer: 2013; pp. 49–56. Kawada, K.; Arima, M.; Miyahara, R.; Tsunomiya, M.; Kikuchi, M.; Yamamoto, F.; Hoshino, A.; Nakajima, Y.; Kinugasa, Y.; Kawano, T. Effect of adding magnifying BLI, magnifying NBI, and iodine staining to white light imaging in diagnosis of early esophageal cancer. Endosc Int Open 2021 , 9 , E1877-E1885, doi:10.1055/a-1583-9196. Tsai, T.J.; Mukundan, A.; Chi, Y.S.; Tsao, Y.M.; Wang, Y.K.; Chen, T.H.; Wu, I.C.; Huang, C.W.; Wang, H.C. Intelligent Identification of Early Esophageal Cancer by Band-Selective Hyperspectral Imaging. Cancers (Basel) 2022 , 14 , doi:10.3390/cancers14174292. Tsai, C.L.; Mukundan, A.; Chung, C.S.; Chen, Y.H.; Wang, Y.K.; Chen, T.H.; Tseng, Y.S.; Huang, C.W.; Wu, I.C.; Wang, H.C. Hyperspectral Imaging Combined with Artificial Intelligence in the Early Detection of Esophageal Cancer. Cancers (Basel) 2021 , 13 , doi:10.3390/cancers13184593. Sohan, M.; Sai Ram, T.; Rami Reddy, C.V. A review on yolov8 and its advancements. In Proceedings of the International Conference on Data Intelligence and Cognitive Informatics, 2024; pp. 529–545. Jiang, P.; Ergu, D.; Liu, F.; Cai, Y.; Ma, B. A Review of Yolo algorithm developments. Procedia computer science 2022 , 199 , 1066–1073. Terven, J.; Córdova-Esparza, D.-M.; Romero-González, J.-A. A Comprehensive Review of YOLO Architectures in Computer Vision: From YOLOv1 to YOLOv8 and YOLO-NAS. Machine Learning and Knowledge Extraction 2023 , 5 , 1680–1716, doi:10.3390/make5040083. Pham, T.-L.; Le, V.-H. Ovarian Tumors Detection and Classification from Ultrasound Images Based on YOLOv8. Journal of Advances in Information Technology 2024 , 15 , 264–275, doi:10.12720/jait.15.2.264-275. Solimani, F.; Cardellicchio, A.; Dimauro, G.; Petrozza, A.; Summerer, S.; Cellini, F.; Renò, V. Optimizing tomato plant phenotyping detection: Boosting YOLOv8 architecture to tackle data complexity. Computers and Electronics in Agriculture 2024 , 218 , doi:10.1016/j.compag.2024.108728. Lakshmi, A.J.; Kishore, G.; Kumar, T.D.; Koushik, V. Drone detection and classification using YOLOv8 and deep CNN. In Proceedings of the International Conference on Innovations and Advances in Cognitive Systems, 2024; pp. 14–27. Ahmed, M.; Afreen, N.; Ahmed, M.; Sameer, M.; Ahamed, J. An inception V3 approach for malware classification using machine learning and transfer learning. International Journal of Intelligent Networks 2023 , 4 , 11–18, doi:10.1016/j.ijin.2022.11.005. Kim, J.-H. Improvement of Inceptionv3 Model Classification Performance Using Chest X-Ray Images. Journal of Mechanics in Medicine and Biology 2022 , 22 , doi:10.1142/s0219519422400322. Mujahid, M.; Rustam, F.; Alvarez, R.; Luis Vidal Mazon, J.; Diez, I.T.; Ashraf, I. Pneumonia Classification from X-ray Images with Inception-V3 and Convolutional Neural Network. Diagnostics (Basel) 2022 , 12 , doi:10.3390/diagnostics12051280. Xia, X.; Xu, C.; Nan, B. Inception-v3 for flower classification. In Proceedings of the 2017 2nd international conference on image, vision and computing (ICIVC), 2017; pp. 783–787. Liu, K.; Yu, S.; Liu, S. An Improved InceptionV3 Network for Obscured Ship Classification in Remote Sensing Images. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 2020 , 13 , 4738–4747, doi:10.1109/jstars.2020.3017676. Muhathir; Dwi Ryandra, M.F.; Syah, R.B.; Khairina, N.; Muliono, R. Convolutional neural network (CNN) of Resnet-50 with Inceptionv3 architecture in classification on X-ray image. In Proceedings of the Computer Science On-line Conference, 2023; pp. 208–221. Z. Riaz, B. Khan, S. Abdullah, S. Khan, and M. S. Islam, “Lung Tumor Image Segmentation from Computer Tomography Images Using MobileNetV2 and Transfer Learning,” Bioengineering , vol. 10, no. 8, p. 981, Aug. 2023, doi: 10.3390/bioengineering10080981. M. S. Islam Sajol, A. S. M Jahid Hasan, M. S. Islam, and M. S. Rahman, “A ConvNeXt V2 Approach to Document Image Analysis: Enhancing High-Accuracy Classification,” in 2024 IEEE 3rd Conference on Information Technology and Data Science (CITDS) , Debrecen, Hungary: IEEE, Aug. 2024, pp. 1–6. doi: 10.1109/CITDS62610.2024.10791343. Nguyen, H. Fast object detection framework based on mobilenetv2 architecture and enhanced feature pyramid. J. Theor. Appl. Inf. Technol 2020 , 98 , 812–824. Dong, K.; Zhou, C.; Ruan, Y.; Li, Y. MobileNetV2 Model for Image Classification. In Proceedings of the 2020 2nd International Conference on Information Technology and Computer Application (ITCA), 2020; pp. 476–480. Hamid, Y.; Wani, S.; Soomro, A.B.; Alwan, A.A.; Gulzar, Y. Smart Seed Classification System based on MobileNetV2 Architecture. In Proceedings of the 2022 2nd International Conference on Computing and Information Technology (ICCIT), 2022; pp. 217–222. Khasoggi, B.; Ermatita, E.; Samsuryadi, S. Efficient mobilenet architecture as image recognition on mobile and embedded devices. Indonesian Journal of Electrical Engineering and Computer Science 2019 , 16 , doi:10.11591/ijeecs.v16.i1.pp389-394. Gulzar, Y. Fruit Image Classification Model Based on MobileNetV2 with Deep Transfer Learning Technique. Sustainability 2023 , 15 , doi:10.3390/su15031906. Additional Declarations No competing interests reported. Supplementary Files supplementary.docx Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 03 Mar, 2026 Reviews received at journal 02 Mar, 2026 Reviews received at journal 01 Mar, 2026 Reviewers agreed at journal 25 Feb, 2026 Reviewers agreed at journal 17 Feb, 2026 Reviewers invited by journal 16 Feb, 2026 Editor assigned by journal 16 Feb, 2026 Editor invited by journal 13 Feb, 2026 Submission checks completed at journal 28 Jan, 2026 First submitted to journal 28 Jan, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8677211","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":592531316,"identity":"dc3290c6-7080-483b-88a8-0153536afc50","order_by":0,"name":"Yu-You Tsai","email":"","orcid":"","institution":"Kaohsiung Armed Forces General Hospital","correspondingAuthor":false,"prefix":"","firstName":"Yu-You","middleName":"","lastName":"Tsai","suffix":""},{"id":592531317,"identity":"01b4c77a-9955-427a-b923-6f97b41fe37b","order_by":1,"name":"Kun-Hua Lee","email":"","orcid":"","institution":"Changhua Christian Hospital","correspondingAuthor":false,"prefix":"","firstName":"Kun-Hua","middleName":"","lastName":"Lee","suffix":""},{"id":592531318,"identity":"820df0a9-1f12-4e0a-96bd-e778065e4695","order_by":2,"name":"Arvind Mukundan","email":"","orcid":"","institution":"National Chung Cheng University","correspondingAuthor":false,"prefix":"","firstName":"Arvind","middleName":"","lastName":"Mukundan","suffix":""},{"id":592531319,"identity":"874bb977-1138-4035-9337-23105f9618b9","order_by":3,"name":"Riya Karmakar","email":"","orcid":"","institution":"National Chung Cheng University","correspondingAuthor":false,"prefix":"","firstName":"Riya","middleName":"","lastName":"Karmakar","suffix":""},{"id":592531320,"identity":"efba285f-4faf-480e-bda5-15dc2ffb891d","order_by":4,"name":"Yaswanth Nagisetti","email":"","orcid":"","institution":"Vel Tech Rangarajan Dr. Sagunthala R\u0026D Institute of Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Yaswanth","middleName":"","lastName":"Nagisetti","suffix":""},{"id":592531321,"identity":"0e4dcb47-0877-45bc-be1a-6f321da4fa32","order_by":5,"name":"Danat Gutema Seyoum","email":"","orcid":"","institution":"Vel Tech Rangarajan Dr. Sagunthala R\u0026D Institute of Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Danat","middleName":"Gutema","lastName":"Seyoum","suffix":""},{"id":592531322,"identity":"881fbb08-3b72-463c-a8d7-6db2e2109810","order_by":6,"name":"Seint Lei Naing","email":"","orcid":"","institution":"Thammasat University","correspondingAuthor":false,"prefix":"","firstName":"Seint","middleName":"Lei","lastName":"Naing","suffix":""},{"id":592531323,"identity":"9e332702-007a-4684-9d52-bd280eb52edf","order_by":7,"name":"Chien-Wei Huang","email":"","orcid":"","institution":"Kaohsiung Armed Forces General Hospital","correspondingAuthor":false,"prefix":"","firstName":"Chien-Wei","middleName":"","lastName":"Huang","suffix":""},{"id":592531324,"identity":"cd840beb-ab54-41b6-8b94-3f513c9b3e01","order_by":8,"name":"Hsiang-Chen Wang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAvElEQVRIiWNgGAWjYBACAwhlA6F4SNCSBtciQayWwyRoMWc/Y/i54Nd5ef4ZCYwP3rYx1BkcIKDFsifHWHpm323DGTcSmA3ntjFIENRicIN3gzRvz+0EA4kENmleoBYzIrRs/s3bcw6khf03sVq2SfP8OAC2hZk4LWfyv1nzNiQbzjjzsFlyzjkJyf0EtRw/lnyb54+dPH978sEPb8ps+CUbCGgBA8Y2MAlSSzhaoOAPsQpHwSgYBaNgRAIAQWY87T4psaQAAAAASUVORK5CYII=","orcid":"","institution":"National Chung Cheng University","correspondingAuthor":true,"prefix":"","firstName":"Hsiang-Chen","middleName":"","lastName":"Wang","suffix":""}],"badges":[],"createdAt":"2026-01-23 09:08:54","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8677211/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8677211/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":103258797,"identity":"32136a19-0c47-4707-8aeb-84c3386dd6a9","added_by":"auto","created_at":"2026-02-23 17:33:36","extension":"jpeg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":651553,"visible":true,"origin":"","legend":"\u003cp\u003eThe process of converting WLI into SAVE\u003c/p\u003e","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8677211/v1/aad3fa88a3a081fed73e6842.jpeg"},{"id":103258796,"identity":"ad0c683b-2c2e-4dfa-b1c3-711df3263467","added_by":"auto","created_at":"2026-02-23 17:33:36","extension":"jpeg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":694646,"visible":true,"origin":"","legend":"\u003cp\u003eThe prediction results of the WLI and SAVE model for the various classes of esophageal cancer. (a) The WLI image of the normal class; (b)The WLI image of the dysplasia class; (c) The WLI image of the SCC class; (d) The SAVE image of the normal class; (e) The SAVE image of the dysplasia class; (f) The SAVE image of the SCC class\u003c/p\u003e","description":"","filename":"floatimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8677211/v1/8360d677b35525b603b02c07.jpeg"},{"id":103506044,"identity":"0fdca917-54e5-4234-bf47-18737ad49817","added_by":"auto","created_at":"2026-02-26 13:33:55","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2207642,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8677211/v1/cf204bf2-d0d2-45a1-b836-5b7cef853187.pdf"},{"id":103258798,"identity":"5b82ce65-4fb2-41ba-aa56-2541290396cf","added_by":"auto","created_at":"2026-02-23 17:33:37","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":16948304,"visible":true,"origin":"","legend":"","description":"","filename":"supplementary.docx","url":"https://assets-eu.researchsquare.com/files/rs-8677211/v1/bf2308200e1029921e6e7478.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Enhancing Esophageal Cancer Detection via Virtual NBI: A Novel Spectrum-Aided Vision Enhancer (SAVE) and Deep Learning Framework","fulltext":[{"header":"Introduction","content":"\u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eGastroenterology is a branch of medicine that studies the digestive system, its disorders, treatment, diagnosis, and management of gastrointestinal bleeding. Gastrointestinal (GI) tract cancer encompasses a variety of cancers which can cause GI bleeding and are life-threatening. Some of them are stomach cancer, esophageal cancer, colon cancer, small intestine cancer, and others [1]. According to recent statistics, esophageal cancer is the seventh most common cancer, with 572,000 new cases annually, and it ranks 6th in mortality with more than 509,000 deaths. The highest regional reports are indicated in Eastern Asia, with Mongolia and China mentioned among the top five worldwide [2]. Squamous cell carcinoma (SCC) and esophageal adenocarcinoma (EAC) are the two main histological subtypes of this male-dominant aggressive malignancy, each indicates different geographical and racial distribution patterns [3]. SCC is the most common subtype, which is mostly reported in the region of Asia and Africa, whereas EAC is more frequently observed in northern America and Europe and is associated with obesity and gastroesophageal reflux disease (GERD) [4]. Epidemiological research has shown that smoking and drinking alcohol are a significant risk factor for SCC.\u003c/p\u003e \u003cp\u003eThe two main abnormalities of esophageal cancer, which are dysplasia, and SCC have their own character and identification. SCC mostly affects the middle third of the esophagus and less frequently it can also affect the lower part. However, the cases are rare in the upper esophagus. Squamous dysplasia is a histological lesion kept within the epithelium, characterized by both architectural and cytologic abnormal growth. Architectural changes are described as the loss of cell polarity and lack of surface maturation. Cytologic abnormalities include enlarged nuclei, hyperchromasia (black stained nuclei), pleomorphism (variation in cell shape and size), and increased or abnormal mitotic activity [5]. Metaplasia is a process in which the normal squamous cell layer of the esophagus is changed to columnar cells.\u003c/p\u003e \u003cp\u003eOver the past decade, artificial intelligence (AI) applications especially machine learning (ML) have achieved excellent advancement in the field of computer vision [6]. The study focused on classifying the early stages of esophageal cancer is utilizing ResNet50 and a support vector machine (SVM) to classify Barrett\u0026rsquo;s esophagus (BE) and esophagitis [7]. In the other study done in 2021, the Inception-v2 model was used to classify GI computed tomography (CT) images to diagnose esophageal cancer [8]. To detect and classify various types of esophageal tissue abnormalities white light imaging has been a very common endoscopic technique used for medical image analysis and showed less accurate results than computer-aided detections CADe [9]. In the comparison study of the diagnostic capacity of convolutional neural networks (CNN) and endoscopists, the result demonstrates that CNN is showing better results in detecting early cancer. CNN was organized by utilizing 13,584 endoscopic images from 2639 lesions of cancer and the result is curated using a test dataset of 2940 images from 140 classes [10]. In the performance comparison of AI versus expert endoscopy, images of 23,746 from 1544 cases of esophageal SCCs and images of 4587 from 458 noncancerous tissues and normal tissues were utilized to build an AI system. Since, the accuracies, specificities, and sensitivities for the identification of SCC were high in percentage, the AI system performed well [11]. The major limitations of WLI shown in these comparisons include being less sensitive to point out the mucosal and vascular changes that are indicative of early or serious stages of cancer, using broad-spectrum light which doesn\u0026rsquo;t use specific light wavelength to identify the lesions clearly, and having a longer learning curve [12]. However, most of the studies use WLI images which utilize three primary color bands, integrating these highly performed AI systems with hyperspectral imaging (HSI) is an innovative way of capturing images beyond WLI and brings high assurance in medical fields. For instance, the five-layered CNN with the dataset of 300 HSI was trained and fine-tuned then it achieved 94.3% accuracy [13,14].\u003c/p\u003e \u003cp\u003eHSI which captures and processes information across the electromagnetic spectrum is an emerging technique for medical imaging [15]. Basically, HSI operates by capturing images based on wavelength variation using spectroscopic techniques. The photons scattering in the blood vessel is recorded to produce spectral images across narrow spectral bands, particularly resulting in multiple bands. For specific bands, the relative amount of light absorbance or reflectance which helps to increase the contrast of cancerous tissues for each image will be done. This information is compiled into a finite 3D volume with two spatial dimensions (X rows and Y columns) and one spectral dimension (λ wavelength), known as a hyperspectral image, hyperspectral (HS) cube, or hypercube [16,17]. The working principle of an HSI that makes use of particular cameras for implementation purposes is divided into main 3 techniques: spectral scanning, spatial scanning, and snapshot method [18]. However, these three techniques are expensive and not easy to use because they require highly specialized cameras such as pushbroom which are more complex than normal imaging systems. They also need high-quality optics for the spectrometer or filters to separate and capture specific wavelengths, and they use precise calibration to ensure accurate data. Hence, the more specific technique called narrow-band imaging (NBI) which is obtained from HSI conversion, helps in getting more convenient and high-resolution images.\u003c/p\u003e \u003cp\u003eNBI is an endoscopic method that makes use of green and blue lights to highlight mucosal and submucosal vasculature [19]. The shorter wavelength of these visible lights (415 nm and 540 nm ) has low tissue permeability which makes it a highly preferable way for identifying mucosal surface structure [20,21]. Because of the strong ability to be absorbed by hemoglobin and its low penetration behavior blue light is used effectively in superficial layers of mucosa. The green light has a slightly longer wavelength and can penetrate a bit deeper than the blue light utilized to visualize the submucosal blood vessel [22]. This NBI method was described in 2003 for the first time and helps to facilitate the detection of early SCC [23]. The study which was done in 1780 esophageal cancer images that include 935 NBI images and 845 WLI, NBI showed better results for the SCC group and good estimation for the normal group in comparison [24]. Similar research was done at the Kaohsiung Medical University for 45 patients and proved that the diagnosis of NBI has greater sensitivity compared with that of white light endoscopy [25].\u003c/p\u003e \u003cp\u003eTherefore, in this work, NBI has been combined with HSI and used to select a few spectrums to enhance the contrast of esophageal cancer identification, particularly dysplasia before it turns into SCC. The CAD system is the best way to increase the visualization of esophageal tissue and enable accurate differentiation among normal, dysplasia, and SCC especially if it integrates with SAVE. By applying the SAVE model and training different ML models for the given dataset, this work focuses on creating a robust model that can identify and prevent pre-cancerous conditions. For the given dataset the captured WLI images are converted into SAVE and made available for further training.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Dataset\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe dataset that is used in this study is sourced from Kaohsiung Armed Forces General Hospital, comprising a total of 5370 images. The given 5370 images are captured using an Olympus endoscope (CV-290, Olympus), and the WLI images are converted into SAVE images to get high performance of the ML models and identify early-stage esophageal cancer. The dataset is comprised of 2 main types of esophageal abnormalities and the normal class for better results. Dysplasia is the earliest stage of esophageal cancer and SCC refers to the serious stage of the tumor. For the model training purpose, the given dataset was split into a train, validation, and test with a total number of images divided into training, validation, and testing sets with a ratio of 70:20:10, respectively. Each split of the dataset contains three main classes and specifically the training dataset contains 3 classes with image distribution of 1420, 1175, and 1160 for the dysplasia, normal, and SCC respectively for both WLI and SAVE images (overall distribution of images per class is summarized in Table S2). All methods were carried out in accordance with relevant guidelines and regulations. The experimental protocols involving human subjects were approved by the Institutional Review Board of Kaohsiung Armed Forces General Hospital (Approval No. KAFGHIRB 114\u0026thinsp;\u0026minus;\u0026thinsp;022). Due to the retrospective nature of the study and the use of de-identified images, the requirement for informed consent was waived by the Institutional Review Board of Kaohsiung Armed Forces General Hospital.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 SAVE\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThis study introduces the Spectrum-Aided Vision Enhancer (SAVE), a novel image processing technique designed to transform standard White Light Imaging (WLI) into hyperspectral-enhanced representations (HSI and SAVE images). The primary objective of SAVE is to augment the diagnostic capability of conventional endoscopy by recovering detailed spectral information that is typically lost in traditional RGB imaging. The core of the SAVE mechanism lies in the precise calibration of endoscopic RGB data to match the reflectance spectra obtained from a spectrometer. To establish a robust mapping between the WLI images and the ground-truth spectral data, we utilized a Macbeth Color Checker (X-Rite Classic), which contains 24 standardized color patches representing natural reflectance spectra. The calibration pipeline begins by converting the raw endoscopic images from the standard sRGB color space into the CIE 1931 XYZ color space, a device-independent standard in colorimetry where X, Y, and Z correspond to the trichromatic response functions. To ensure high-fidelity color reproduction, we applied a series of correction algorithms to address common imaging artifacts, including non-linear sensor response, dark current noise, imperfect color separation, and lens distortion. The corrected XYZ values (XYZ\u003csub\u003ecorrect\u003c/sub\u003e) are derived through this rigorous calibration process. The transformation from raw RGB to the initial XYZ space is formulated in Eq.\u0026nbsp;(\u003cspan refid=\"Equ1\" class=\"InternalRef\"\u003e1\u003c/span\u003e), while the subsequent error correction and calculation of the final XYZ\u003csub\u003ecorrect\u003c/sub\u003e are defined in Eq.\u0026nbsp;(\u003cspan refid=\"Equ2\" class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Equ1\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$\\:\\left[C\\right]=\\left[XY{Z}_{Spectrum}\\right]\\times\\:pinv\\left(\\left[V\\right]\\right)$$\u003c/div\u003e \u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e \u003cdiv id=\"Equ2\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e\n$$\\:\\left[XY{Z}_{Correct}\\right]=\\left[C\\right]\\times\\:\\left[V\\right]$$\u003c/div\u003e \u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFollowing the initial color space conversion, the calibration process necessitates a rigorous alignment between the generated SAVE representations and the clinical gold standard\u0026mdash;Narrow Band Imaging (NBI) captured by Olympus endoscopes. To achieve this, we employed Principal Component Analysis (PCA) to extract the most significant features from the high-dimensional reflectance spectra. This step is crucial for reducing data dimensionality while preserving essential spectral information, thereby optimizing computational efficiency. Our PCA analysis revealed that the first six principal components were sufficient to capture 99.64% of the total variance in the spectral data. This finding confirms that a low-dimensional representation can effectively model the complex reflectance properties of esophageal tissue. To validate the accuracy of our color calibration, we calculated the Root Mean Square Error (RMSE) between the corrected image values (XYZ\u003csub\u003ecorrect\u003c/sub\u003e) and the ground-truth spectrometer measurements (XYZ\u003csub\u003espectrum\u003c/sub\u003e). The resulting mean RMSE of 0.19 indicates a minimal color difference, demonstrating the high fidelity of our calibration pipeline.\u003c/p\u003e \u003cp\u003eThe successful calibration serves as the foundation for converting standard WLI images into Hyperspectral Imaging (HSI) formats and subsequently into virtual NBI (SAVE) images. To formalize this conversion, we conducted a multiple regression analysis to model the relationship between the input WLI data and the target spectral components. This analysis yielded a transformation matrix, defined in Eq.\u0026nbsp;(\u003cspan refid=\"Equ3\" class=\"InternalRef\"\u003e3\u003c/span\u003e), which serves as the mathematical kernel for mapping the broad-spectrum WLI signals to the specific narrow-band wavelengths required for SAVE. By applying this matrix, we can synthetically reconstruct tissue visualization that mimics the contrast-enhancement properties of hardware-based NBI, effectively highlighting vascular structures and mucosal patterns.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Equ3\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ3\" name=\"EquationSource\"\u003e\n$$\\:\\left[M\\right]=\\left[Score\\right]\\times\\:pinv\\left(\\left[{V}_{Color}\\right]\\right)$$\u003c/div\u003e \u003cdiv class=\"EquationNumber\"\u003e3\u003c/div\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe analog spectrum derived from the corrected XYZ values Equation. \u003cspan refid=\"Equ4\" class=\"InternalRef\"\u003e4\u003c/span\u003e demonstrated minimal color discrepancies compared to real measurements, supporting the system\u0026rsquo;s accuracy:\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Equ4\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ4\" name=\"EquationSource\"\u003e\n$$\\:{\\left[SSpectrum\\right]}_{380\\sim780nm}=\\left[EV\\right]\\left[M\\right]\\left[{V}_{Color}\\right]$$\u003c/div\u003e \u003cdiv class=\"EquationNumber\"\u003e4\u003c/div\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eUpon reconstructing the hyperspectral data from RGB inputs, the subsequent phase involved synthesizing virtual NBI representations and validating their fidelity against clinical standards. We utilized the Olympus NBI system as the ground truth benchmark, comparing its output directly with our algorithmically generated SAVE images. This comparison was facilitated using the 24-patch color checker to ensure consistent colorimetric evaluation. To quantify the perceptual similarity between the synthetic SAVE images and the hardware-captured NBI images, we computed the CIEDE2000 color difference metric (ΔE\u003csub\u003e00\u003c/sub\u003e) for each color patch. The initial analysis yielded a mean ΔE\u003csub\u003e00\u003c/sub\u003e of 2.79, indicating a high degree of visual resemblance with only minor deviations. These discrepancies were primarily attributed to inherent differences in the illumination spectra, reflectance properties, and color-matching functions between the simulated environment and the physical endoscopic hardware. A critical observation was that the spectral mismatch was most pronounced in the 450\u0026ndash;540 nm wavelength range, which is vital for visualizing superficial mucosal capillaries. To mitigate this and further minimize the color divergence, we implemented a secondary calibration layer utilizing the Cauchy-Lorentz distribution. This statistical model, defined in Eq.\u0026nbsp;(\u003cspan refid=\"Equ5\" class=\"InternalRef\"\u003e5\u003c/span\u003e), is particularly effective for characterizing spectral resonance and was applied to fine-tune the system's spectral response. By adjusting the distribution parameters, we significantly reduced the spectral deviation, ensuring that the SAVE images provided a clinically accurate emulation of the NBI modality.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Equ5\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ5\" name=\"EquationSource\"\u003e\n$$\\:f\\left(x;{x}_{0},\\gamma\\:\\right)=\\frac{1}{\\pi\\:\\gamma\\:\\left[1+{\\frac{\\left(x-{x}_{0}\\right)}{\\gamma\\:}}^{2}\\right]}=\\frac{1}{\\pi\\:}\\left[\\frac{\\gamma\\:}{\\left(x-{x}_{0}\\right)2+{\\gamma\\:}^{2}}\\right]$$\u003c/div\u003e \u003cdiv class=\"EquationNumber\"\u003e5\u003c/div\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eTo refine the light spectrum parameters, we employed the dual annealing optimization algorithm. This stochastic approach integrates the robustness of Classical Simulated Annealing (CSA) with the efficiency of Fast Simulated Annealing (FSA), allowing for a global search of the optimal spectral configuration. Following this optimization, the mean CIEDE2000 color difference was stabilized at 5.36. While this value is slightly higher than the initial calibration, it accounts for the complex post-processing inherent in the Olympus system\u0026mdash;specifically, the introduction of brownish tones alongside the primary hemoglobin absorption peaks (415 nm and 540 nm) to enhance tissue realism. To rigorously assess the fidelity of the SAVE-generated images against the gold-standard NBI, we utilized a triad of image quality metrics: Structural Similarity Index Measure (SSIM), Entropy, and Peak Signal-to-Noise Ratio (PSNR). The SSIM analysis yielded a similarity score of 94.27%. This high correlation confirms that the SAVE mechanism successfully preserves the structural details and textural information of the mucosal surface, making it a reliable surrogate for hardware NBI. Entropy was calculated to evaluate the richness of image information. The results showed a marginal difference of only 0.37% between the SAVE and Olympus NBI images, indicating that our method maintains the textural complexity required for diagnosis. The PSNR, a standard metric for reconstruction quality, reached 27.88 dB. This value suggests that the SAVE images retain high signal fidelity with minimal noise introduction during the transformation process. The comprehensive workflow, from initial RGB input to the final optimized SAVE output, is schematically illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 DL algorithm\u003c/h2\u003e \u003cdiv id=\"Sec6\" class=\"Section3\"\u003e \u003ch2\u003e2.3.1 YOLOv8\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe first YOLO model was introduced in 2015 by Joseph Redmon in a C-based repository named Darknet during his Ph.D. studies at the University of Washington. Since then, the community has continued developing and enhancing the subsequent YOLO version [26,27]. Ultralytics the community that introduced YOLOv5 released YOLOv8 in January 2023. This new version supports different computer vision tasks including object detection, object segmentation, tracking, and classification. By targeting the specific need of the task YOLOv8 developed with five different variants including YOLOv8n (nano), YOLOv8s (small), YOLOv8m (medium), YOLOv8l (large), and YOLOv8x (extra-large) [28,29].\u003c/p\u003e \u003cp\u003eThe backbone responsible for the feature extraction of the input images is one of the main components of the YOLOv8 architecture. By making use of the path aggregation network (PANet) with feature pyramid network (FPN) the other architecture called neck detects objects at various levels(an illustration of the YOLOv8 architecture used in this study is shown in Figure \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e). The detection head of YOLOv8 is where the final prediction is done and for tasks like object classification and regression tasks decoupled head is utilized [30,31]. For different tasks, YOLOv8 uses different Loss functions. The classification loss utilizes Variational Focal Loss (VFL), and the regression employs a combination of Distribution Focal Loss (DFL) and Complete IoU (CIoU) Loss. Unlike Focal Loss (FL) and Quality Focal Loss (QFL), which are symmetric, VFL uses an asymmetric weighting technique to find out the solution of the imbalance between negative and positive samples.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Equ6\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ6\" name=\"EquationSource\"\u003e\n$$\\:VF{L}_{\\left(p,q\\right)}=\\left\\{-q\\left(q\\:\\text{log}\\text{log}\\left(P\\right)+\\left(1-q\\right)\\text{log}\\left(1-p\\right)\\right),q\u0026gt;0-\\alpha\\:{P}^{y}\\text{log}\\left(1-p\\right),q=0\\right\\}$$\u003c/div\u003e \u003cdiv class=\"EquationNumber\"\u003e6\u003c/div\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eWhere VFL stands for variational focal loss, p represents the label and q is a value calculated by normalized alignment matric if a positive sample is taken.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section3\"\u003e \u003ch2\u003e2.3.2 InceptionV3\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eTransfer learning is a way of solving a new problem based on the previous learning history of the model. The main characteristic of the neural network is that the layer that is closer to the output is updated based on the new data, while the remaining hidden layers are kept unchanged [32]. InceptionV3 is one of the transfer learning models built as an advanced version of Google Net (InceptionV1) which achieved the first prize in the 2014 ILSVRC Competition [33]. This model is the other version of Inception-V1 and Inception-V2 architectures and it\u0026rsquo;s deep CNN which is trained on low-configuration computers [34]. Moreover, the model is trained on the large-scale ImageNet dataset and has the ability to classify images into 1000 different groups [35]. The architecture of inceptionV3 can be divided into smaller convolutional kernels which has a huge effect of lowering the model\u0026rsquo;s parameter and minimizing the probability of overfitting [36]. The main architecture includes convolution, MaxPool, Average Pool, fully connected, dropout, and SoftMax [37]. The loss function of the inceptionV3 for the classification task is represented in Eq.\u0026nbsp;\u003cspan refid=\"Equ7\" class=\"InternalRef\"\u003e7\u003c/span\u003e.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Equ7\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ7\" name=\"EquationSource\"\u003e\n$$\\:L=-\\:{\\sum\\:}_{i=1}^{c}{y}_{i}\\text{log}\\left({x}_{i}\\right)$$\u003c/div\u003e \u003cdiv class=\"EquationNumber\"\u003e7\u003c/div\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eWhere y\u003csub\u003ei\u003c/sub\u003e is the true value of class i, x\u003csup\u003ei\u003c/sup\u003e is the predicted probability for class I, and C is the overall number of classes.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section3\"\u003e \u003ch2\u003e2.3.3 ConvNeXt-V2\u003c/h2\u003e \u003cp\u003eConvNeXt-V2 is a modern convolutional neural network model that aims at closing the performance gap between traditional CNN models and models based on transformers. It uses the concepts of the self-attention mechanism of transformers by a Fully Convolutional Masked Autoencoder (FCMAE) framework, which allows the model to learn more expressive and strong image features. Under this framework, portions of the input image are masked, and the network learns to replicate the missing portions, which enhances its feature-learning capacity. To facilitate this process in ConvNeXt-V2, regular convolutions in the masked regions are substituted with sparse convolutions, and the reconstruction loss of the pre-training is based on Mean Squared Error (MSE). Another major advancement in ConvNeXt-V2 is the integration of Global Response Normalization (GRN), a mechanism designed to mitigate feature collapse\u0026mdash;a phenomenon where deep models saturate or deactivate feature maps, limiting their representational capacity. GRN enhances channel-wise feature competition by normalizing the global activation strength, thereby stabilizing the learning process. Mathematically, for a given input feature map with defined height, width, and channel dimensions, the global response is first computed by aggregating the spatial information across each channel using the L2-norm, as defined in Eq.\u0026nbsp;(\u003cspan refid=\"Equ8\" class=\"InternalRef\"\u003e8\u003c/span\u003e) [38]:\u003cdiv id=\"Equ8\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ8\" name=\"EquationSource\"\u003e\n$$\\:G\\left(X\\right)=\\left\\{\\left|\\left|{X}_{1}\\right|\\right|,\\:\\left|\\left|{X}_{2}\\right|\\right|,\\dots\\:,\\left|\\left|{X}_{c}\\right|\\right|,\\right\\}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e8\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eSubsequently, a channel-wise normalization step is applied to compute the relative importance of each channel, denoted as N(X), as shown in Eq.\u0026nbsp;(\u003cspan refid=\"Equ9\" class=\"InternalRef\"\u003e9\u003c/span\u003e):\u003cdiv id=\"Equ9\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ9\" name=\"EquationSource\"\u003e\n$$\\:N\\left(\\left|\\left|{X}_{i}\\right|\\right|\\right)=\\frac{\\left|\\left|{X}_{i}\\right|\\right|}{\\sum\\:_{j=1}^{C}\\left|\\left|{X}_{j}\\right|\\right|}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e9\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eFinally, the calibrated feature map is obtained by scaling the original input with learnable scaling and shifting parameters, ensuring that the network retains the flexibility to adaptively emphasize informative features. This transformation is expressed in Eq.\u0026nbsp;(\u003cspan refid=\"Equ10\" class=\"InternalRef\"\u003e10\u003c/span\u003e):\u003cdiv id=\"Equ10\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ10\" name=\"EquationSource\"\u003e\n$$\\:{X}_{i}^{{\\prime\\:}}=\\gamma\\:\\bullet\\:{X}_{i}\\bullet\\:N\\left(G\\left({X}_{i}\\right)\\right)+\\beta\\:+{X}_{i}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e10\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eWhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\gamma\\:\\)\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\beta\\:\\)\u003c/span\u003e\u003c/span\u003e are learnable parameters\u003c/p\u003e \u003cp\u003eConvNeXt-V2 was employed in this work to classify images of the esophagus (Normal, Dysplasia, and SCC) with the help of WLI and SAVE-generated hyperspectral band-selected data. Resizing was done using all the images to the size of 224x 224 pixels prior to training. Due to the design of ConvNeXt-V2, focusing on high-accuracy classification with high generalization, the algorithm is compatible with medical images where the slightest changes in tissue texture and color can need to be detected. In order to guarantee the dependability of the assessment, 5-fold cross-validation was used, and the average performance scores across all folds were obtained. This enabled the model to be tested equally on WLI and the hyperspectral inputs.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section3\"\u003e \u003ch2\u003e2.3.4 Inception-ResNet-V2\u003c/h2\u003e \u003cp\u003eInception-ResNet-V2 is a deep convolutional neural network, which combines the multi-scale feature extracting capability of Inception modules with the stability of residual learning. The model starts with a stem block which is capable of basic feature extraction and then there are three major blocks, Inception-ResNet-A, Inception-ResNet-B, and Inception-ResNet-C. Many convolutional branches operating in parallel are contained in each block, and this aids the network to acquire features of images at various scales.\u003c/p\u003e \u003cp\u003eA defining characteristic of this architecture is the incorporation of shortcut connections (skip connections) within each Inception block. These connections compel the network to learn a residual mapping rather than a direct transformation, which significantly enhances training stability and mitigates the vanishing gradient problem in deep networks. The residual learning mechanism is formally defined in Eq.\u0026nbsp;(\u003cspan refid=\"Equ11\" class=\"InternalRef\"\u003e11\u003c/span\u003e) [39]:\u003cdiv id=\"Equ11\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ11\" name=\"EquationSource\"\u003e\n$$\\:H\\left(x\\right)=F\\left(x\\right)+x$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e11\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eWhere \u0026#119909; represents the input to the block, whereas \u0026#119865;(\u0026#119909;) is the multi-branch convolutional output. All the blocks in the architecture are done with batch normalization and reduction in order to accelerate training and regulate model complexity. The Inception-ResNet-V2 has a number of hundreds of layers, which enables it to make both fine and complex contributions to medical images. The model was applied in this research to categorize esophageal images of the WLI and SAVE. All the images were downscaled to 224 x 224 pixels and then submitted to the network. To make the evaluation reliable, a 5-fold cross-validation procedure was performed, i.e. the dataset was split into five equal parts and model was trained and validated five times. This increased the performance comparison between the WLI and the hyperspectral images in terms of being more accurate and unbiased.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section3\"\u003e \u003ch2\u003e2.3.5 MobileNetV2\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eMobileNetV2 is one of the neural networks which is developed by Google and it runs on any device efficiently with less computational power and gives high accuracy [40]. The architecture of the model contains an inverted residual block, linear bottleneck layer, depth-wise separable convolution, ReLU6 Activation, expansion layer (1X1 convolution), stridden convolutions for down-sampling, global average pooling (GAP), and final fully connected layer [41,42]. The model depends on depth-wise separable convolutions which separate standard convolution into two steps: depth-wise (pre-channel filtering) and pointwise (combining information across channels). This is important to reduce the computational complexity compared to traditional convolutions while maintaining accuracy [43]. The other architecture of mobilenetv2 which are inverted residual and linear bottleneck structures significantly reduce the computational cost of convolutions making it both efficient and effective to use [44].\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePerformance of various ML models for a given dataset.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModels\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eImaging Modalities\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eClasses\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eF1 Score\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSpecificity\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"5\" rowspan=\"6\"\u003e \u003cp\u003e\u003cb\u003eYOLOv8n\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eWLI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNormal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e85\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDysplasia\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e93\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e97\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSCC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e97\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eSAVE\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNormal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e85\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDysplasia\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e96\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSCC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e81\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e94\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"5\" rowspan=\"6\"\u003e \u003cp\u003e\u003cb\u003eInceptionV3\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eWLI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNormal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e93\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e89\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDysplasia\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e96\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSCC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e98\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eSAVE\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNormal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e95\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e92\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDysplasia\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e96\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSCC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e100\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e100\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"5\" rowspan=\"6\"\u003e \u003cp\u003e\u003cb\u003eInception-ResNet-V2\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eWLI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNormal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e93\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e94\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDysplasia\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e94\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSCC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e97\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eSAVE\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNormal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e98\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDysplasia\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e93\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSCC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e96\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"5\" rowspan=\"6\"\u003e \u003cp\u003e\u003cb\u003eConvNeXt-V2\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eWLI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNormal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e85\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDysplasia\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e96\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e74\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e98\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSCC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e93\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e94\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eSAVE\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNormal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e93\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e96\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDysplasia\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e94\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSCC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e96\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"5\" rowspan=\"6\"\u003e \u003cp\u003e\u003cb\u003eMobileNetV2\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eWLI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNormal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e95\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e89\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDysplasia\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e88\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSCC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e89\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eSAVE\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNormal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e88\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDysplasia\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e96\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSCC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e100\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e100\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe cross-validation that was performed in five-fold in all five deep learning models, i.e., YOLOv8n, InceptionV3, Inception-ResNet-V2, ConvNeXt-V2, and MobileNetV2, demonstrated different trends in the diagnostic performance of the models with application to WLI and SAVE imaging modalities. This variation in precision, recall, F1-score, and specificity between the Normal, Dysplasia as well as SCC classes can be directly explained by the variation in model architecture and the quality of visual information that each imaging modality possesses. SAVE scans always provided better clarity, contrast, and structural visibility, which contributed greatly to the change of the pattern of performance in terms of classes. Fold-wise numerical results are given in Supplementary Tables S1-S25 and all fold-wise loss plots and confusion matrices are found in Figures \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e-S90.\u003c/p\u003e \u003cp\u003eThe first one began with YOLOv8n whose lightweight structure was relatively sensitive to the gains made by SAVE imaging. There was a minor improvement in dysplasia performance under SAVE mainly because of the increase in the level of illumination and clarity of the epithelial pattern by which early-stage morphological variations, which are weak in WLI, can be detected. This has led to more increase in recall demonstrating that a smaller number of dysplastic cases were overlooked. But SCC recall declined with SAVE as greater strengthening of intermediate dysplastic features by the modality introduced minor overlaps between high-grade dysplasia and early SCC leading to misclassification at the border between the two classes. Normal class was also similar in modalities as YOLOv8n already works well with normal tissue and the extra SAVE clarity did not dramatically change the property of the class. These results indicate the consistency of YOLOv8n and also exemplify its weaknesses in discriminating closely related pathological classes in the case of adding visual complexity through SAVE. Fold-wise numerical results are given in Supplementary Tables S11-S15 and all fold-wise loss plots and confusion matrices are found in Figures S71-S90.\u003c/p\u003e \u003cp\u003eInceptionV3, on the contrary, was more advantaged by SAVE images. Its multi-scale convolutional filters are based heavily on rich spatial detail and color gradient both of which are more exhibited in SAVE. The recall of dysplasia (strong) and near-perfect recall showed an improvement of values under SAVE, which highlights the fact that the latter is more accurate in the detection of small lesions. Equally, precision in SCC was 100% with SAVE, showing that more evident morphological indicators resulted in the model being able to remove false positives completely. The sharpening of the epithelial edges, and the even distribution of light in SAVE also led to improved normal tissue recognition since the model could avoid misclassifying benign abnormalities that were a common occurrence in WLI. Collectively, these improvements affirm that the InceptionV3 architecture is extremely compatible with the sophisticated spectral and structural data delivered by SAVE which allows to perform more confident classification in all tissue types. Fold-wise numerical results are given in Supplementary Tables S1-S5 and all fold-wise loss plots and confusion matrices are in Figures S21-S40.\u003c/p\u003e \u003cp\u003eInception-ResNet-V2 also produced good and robust performance in both modalities, however, SAVE provided some additional benefits that enhanced the reliability of the classification. This architecture has residual connections together with inception modules which makes it especially effective at refining intermediate features. The highest improvement was observed with the Normal class with SAVE whereby the precision was enhanced significantly. This was the case because SAVE lowers the appearance of shadows, glare and minor inflammatory alterations which may appear like pathology and enhances the confidence of the model to recognize non-pathological mucosa. Dysplasia performance was also very strong but the precision was very slight because SAVE tended to exaggerate benign textural changes which made the model over-detect dysplasia in normal areas. It was observed that SCC performance was also balanced, with both precision and recall improving or retaining high levels due to the fact that SCC morphology is refined obviously in SAVE and with ease in the deep layers of the feature extraction. The findings highlight the versatility of the Inception-ResNet-V2 that successfully used the better visual cues of SAVE to sharpen the predictions without causing instability. Fold-wise numerical results are given in Supplementary Tables S16-S20 and all fold-wise loss plots and confusion matrices are in Figures S61-S70.\u003c/p\u003e \u003cp\u003eConvNeXt-V2 model showed the most drastic improvements with SAVE and particularly with dysplasia class. This model is based on the transformer-style architectural ideas and relies on the detailed texture patterns and long-range dependencies which are better offered by SAVE imaging than by WLI. Recall of dysplasia increased significantly when under WLI, it was low and when under SAVE, it was almost perfect, which is that ConvNeXt-V2 needs high-quality structural and textural detail in order to identify early or mild lesions. Normal and SCC classes were also benefitted and displayed more equalized values of precision and recall with SAVE. WLI weaknesses of irregular brightness, contrast, and lesser expression of gentle architectural distortion were barriers to the generalizability of this architecture in some classes. SAVE addressed these errors and enabled ConvNeXt-V2 to see the tissue patterns in a more holistic and correct way. These findings prove that highly vision transformer-like models are best used in cases of improved imaging conditions. Fold-wise numerical results are given in Supplementary Tables S21-S25 and all fold-wise loss plots and confusion matrices are in Figures \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e-S20.\u003c/p\u003e \u003cp\u003eLastly, even though MobileNetV2 is lightweight, it also showed significant gains with SAVE. This model suffers a shortcoming of not being able to extract features of WLI limited features of distinguishing dysplasia and borderline pathology. To counter this, SAVE made available more explicit micro-structural characteristics that made the boundaries of decisions easier to find in the model. The Dysplasia F1-scores were better and SCC precision was 100% indicating that MobileNetV2 was able to reliably discriminate between advanced malignant appearances when assisted by high-quality imaging. There was also the improvement of the Normal class as SAVE reduced noise, asymmetric lighting, and other minor mucosal artifacts that previously led to a false positive. Therefore, MobileNetV2 has shown that also small models can significantly benefit the added structural and spectral cues of SAVE that can be used to achieve performance in some classes similar to much larger designs. Fold-wise numerical results are given in Supplementary Tables S6-S10 and all fold-wise loss plots and confusion matrices are found in Figures S41-S60.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe result of this study demonstrates that the SAVE model has better results than the conventional one. Hence by enhancing the early identification of esophageal cancer, the SAVE model can contribute to better patient outcomes, as identifying and treating early cancer can prevent the spread of cancer into the severe stage. Kaohsiung Medical University Chung-Ho Memorial Hospital used as the only source of the dataset which could limit the model\u0026rsquo;s generalization. Incorporating data from different sources like multiple hospitals, different regions, races, and age groups would help to consider the geographical and genetic variability of esophageal cancer and races could increase the model's capacity to find esophageal cancer across various demographic groups.\u003c/p\u003e \u003cp\u003eEven though the scope of the dataset is somewhat limited to identifying and diagnosing all the types of esophageal cancer, this study aimed to establish a benchmark for ML models applied to esophageal cancer detection. Having research done on SCC which is the main type of esophageal cancer type is very crucial, as it represents both ends of the spectrum early, hard-to-detect cases (dysplasia), and severe, late-stage cases of the disease (SCC). the other important types of esophageal cancer that are not mentioned in this work include Barrett's esophagus and high- and -low-grade dysplasia. Barrett's esophagus is a stage in which the normal lining of the esophagus is replaced with the other abnormal tissue and this is the result of long-term acid reflux most of the time. the word low-grade dysplasia refers to starting early abnormal growth of the esophagus lining but it doesn't directly threaten human life and has a chance of being treated and cured. the highest stage of low-grade dysplasia is called high-grade dysplasia. the inclusion of these different varieties of cancer in the study and testing the performance of the SAVE model toward the WLI is the future direction of this research.\u003c/p\u003e \u003cp\u003eFor comparing the performance of the SAVE model with the WLI model and analyzing different architecture's performance in the identification of different types of esophageal cancer, five different ML models were utilized. this comparative technique makes the current study a great benchmark for future studies that may include many more ML models and a larger amount of dataset. the identification of which ML model is more effective in medical image tasking and which model needs further fine-tuning or optimization is done by current analysis. the name of the models which was trained for a given dataset includes, YOLOv8, InceptionV3, ConvNeXt-V2, Inception-ResNet-V2, and MobileNetV2. Out of all the tested models, YOLOv8 showed the best result in comparison with the other four models on the basis of accuracy and reliability. Moreover, this study focuses on sensitivity (recall), precision, and f1-score, future work will incorporate ROC curves analysis to better examine the balance between sensitivity and specificity and to provide AUC-based performance metrics. By addressing this the applicability and reliability of the SAVE model could future improve in the medical sector and lead to accurate identification of esophageal cancer.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e"},{"header":"Conclusions","content":"\u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eIn this study, we proposed and validated a novel image enhancement algorithm, the Spectrum-Aided Vision Enhancer (SAVE), designed to overcome the limitations of standard White Light Imaging (WLI) in esophageal cancer detection. By leveraging hyperspectral principles to generate virtual Narrow Band Imaging (NBI) representations, the SAVE mechanism significantly improves the visibility of mucosal vasculature and subtle lesions without the need for specialized hardware. This software-driven approach provides a cost-effective solution for enhancing image quality, which is a critical preprocessing step for automated medical diagnosis. Our comprehensive evaluation using five state-of-the-art deep learning models\u0026mdash;YOLOv8, InceptionV3, Inception-ResNet-V2, ConvNeXt-V2, and MobileNetV2\u0026mdash;demonstrates that the integration of SAVE consistently outperforms traditional WLI-based training. The models trained on SAVE-enhanced datasets exhibited superior accuracy, sensitivity, and specificity in classifying and detecting Squamous Cell Carcinoma (SCC) and dysplasia. These findings confirm that domain-specific image enhancement can effectively augment the feature extraction capabilities of deep neural networks, leading to more robust and reliable Computer-Aided Diagnosis (CAD) systems. Future work will focus on expanding the dataset to include a wider variety of esophageal pathologies and conducting multi-center clinical trials to further validate the generalizability of the SAVE algorithm. Additionally, we aim to optimize the computational efficiency of the SAVE mechanism for real-time deployment in clinical endoscopic systems. Ultimately, this study highlights the potential of combining advanced image processing techniques with deep learning to assist endoscopists in early cancer detection, potentially reducing missed diagnoses and improving patient outcomes.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e \u003ch2\u003eInstitutional Review Board Statement\u003c/h2\u003e \u003cp\u003eThe study was conducted in accordance with the Declaration of Helsinki and approved by the Institutional Review Board of Kaohsiung Armed Forces General Hospital (Protocol code KAFGHIRB 114\u0026thinsp;\u0026minus;\u0026thinsp;022; approved on 25 April 2025). The requirement for informed consent was waived by the Institutional Review Board due to the retrospective design of the study.\u003c/p\u003e \u003c/p\u003e\u003cp\u003e \u003ch2\u003eConflicts of Interest:\u003c/h2\u003e \u003cp\u003eThe authors declare no conflicts of interest.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eFunding:\u003c/h2\u003e \u003cp\u003eThis research received support from the National Science and Technology Council, Republic of China through the following grants: NSTC 113-2221-E-194-011-MY3. Additionally, financial support was provided by the Dalin Tzu Chi Hospital, Buddhist Tzu Chi Medical Foundation-National Chung Cheng University Joint Research Program and Kaohsiung Armed Forces General Hospital Research Program KAFGH_D_115-005 in Taiwan.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eConceptualization **:** R.K., Y.-Y.T., Y.N, A.M. \u0026amp; H.-C.W. ; data curation : Y.-Y.T., D.G.S, K.-H.L.., R.K., A.M. \u0026amp; H.-C.W. ; formal analysis: D.G.S, K.-H.L., Y.N, A.M., H.-C.W. ; funding acquisition Y.-Y.T., K.-H.L., A.M. \u0026amp; H.-C.W. ; investigation: D.G.S, R.K., K.-H.L.; methodology: D.G.S, R.K., C.-Y.K, \u0026amp;amp; A.M. ; project administration: C.-W.H, A.M. and H.-C.W.; resources, C.-W.H, A.M. \u0026amp;amp; H.-C.W.; software: R.K. and C.-W.H. ; supervision, Y.N, H.-C.W.; writing\u0026mdash;original draft, Y.-Y.T., S.L.N. and D.G; writing\u0026mdash;review and editing: R.K., S.L.N., A.M. \u0026amp;amp; H.-C.W. All authors have read and agreed to the published version of the manuscript.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe data presented in this study are available in this article; further considerable requests can be made to the corresponding author (H.-C.W.).\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eDisclaimer/Publisher\u0026rsquo;s Note\u003c/strong\u003e: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eFan, S.; Xu, L.; Fan, Y.; Wei, K.; Li, L. Computer-aided detection of small intestinal ulcer and erosion in wireless capsule endoscopy images. \u003cem\u003ePhys Med Biol\u003c/em\u003e \u003cstrong\u003e2018\u003c/strong\u003e, \u003cem\u003e63\u003c/em\u003e, 165001, doi:10.1088/1361-6560/aad51c.\u003c/li\u003e\n\u003cli\u003eHe, Y.; Li, D.; Shan, B.; Liang, D.; Shi, J.; Chen, W.; He, J. Incidence and mortality of esophagus cancer in China, 2008–2012. \u003cem\u003eChin J Cancer Res\u003c/em\u003e \u003cstrong\u003e2019\u003c/strong\u003e, \u003cem\u003e31\u003c/em\u003e, 426–434, doi:10.21147/j.issn.1000-9604.2019.03.04.\u003c/li\u003e\n\u003cli\u003eAbbas, G.; Krasna, M. Overview of esophageal cancer. \u003cem\u003eAnn Cardiothorac Surg\u003c/em\u003e \u003cstrong\u003e2017\u003c/strong\u003e, \u003cem\u003e6\u003c/em\u003e, 131–136, doi:10.21037/acs.2017.03.03.\u003c/li\u003e\n\u003cli\u003eKojima, T.; Shah, M.A.; Muro, K.; Francois, E.; Adenis, A.; Hsu, C.H.; Doi, T.; Moriwaki, T.; Kim, S.B.; Lee, S.H.; et al. Randomized Phase III KEYNOTE-181 Study of Pembrolizumab Versus Chemotherapy in Advanced Esophageal Cancer. \u003cem\u003eJ Clin Oncol\u003c/em\u003e \u003cstrong\u003e2020\u003c/strong\u003e, \u003cem\u003e38\u003c/em\u003e, 4138–4148, doi:10.1200/JCO.20.01888.\u003c/li\u003e\n\u003cli\u003eJain, S.; Dhingra, S. Pathology of esophageal cancer and Barrett's esophagus. \u003cem\u003eAnn Cardiothorac Surg\u003c/em\u003e \u003cstrong\u003e2017\u003c/strong\u003e, \u003cem\u003e6\u003c/em\u003e, 99–109, doi:10.21037/acs.2017.03.06.\u003c/li\u003e\n\u003cli\u003eBarragan-Montero, A.; Javaid, U.; Valdes, G.; Nguyen, D.; Desbordes, P.; Macq, B.; Willems, S.; Vandewinckele, L.; Holmstrom, M.; Lofman, F.; et al. Artificial intelligence and machine learning for medical imaging: A technology review. \u003cem\u003ePhys Med\u003c/em\u003e \u003cstrong\u003e2021\u003c/strong\u003e, \u003cem\u003e83\u003c/em\u003e, 242–256, doi:10.1016/j.ejmp.2021.04.016.\u003c/li\u003e\n\u003cli\u003eKumar A, C.; Mubarak, D.M.N. Classification of Early Stages of Esophageal Cancer Using Transfer Learning. \u003cem\u003eIrbm\u003c/em\u003e \u003cstrong\u003e2022\u003c/strong\u003e, \u003cem\u003e43\u003c/em\u003e, 251–258, doi:10.1016/j.irbm.2021.10.003.\u003c/li\u003e\n\u003cli\u003eChen, K.B.; Xuan, Y.; Lin, A.J.; Guo, S.H. Esophageal cancer detection based on classification of gastrointestinal CT images using improved Faster RCNN. \u003cem\u003eComput Methods Programs Biomed\u003c/em\u003e \u003cstrong\u003e2021\u003c/strong\u003e, \u003cem\u003e207\u003c/em\u003e, 106172, doi:10.1016/j.cmpb.2021.106172.\u003c/li\u003e\n\u003cli\u003eWang, P.; Liu, P.; Brown, J.R.G.; Berzin, T.M.; Zhou, G.; Lei, S.; Liu, X.; Li, L.; Xiao, X. Lower adenoma miss rate of computer-aided detection-assisted colonoscopy vs routine white-light colonoscopy in a prospective tandem study. \u003cem\u003eGastroenterology\u003c/em\u003e \u003cstrong\u003e2020\u003c/strong\u003e, \u003cem\u003e159\u003c/em\u003e, 1252–1261. e1255.\u003c/li\u003e\n\u003cli\u003eIkenoyama, Y.; Hirasawa, T.; Ishioka, M.; Namikawa, K.; Yoshimizu, S.; Horiuchi, Y.; Ishiyama, A.; Yoshio, T.; Tsuchida, T.; Takeuchi, Y.; et al. Detecting early gastric cancer: Comparison between the diagnostic ability of convolutional neural networks and endoscopists. \u003cem\u003eDig Endosc\u003c/em\u003e \u003cstrong\u003e2021\u003c/strong\u003e, \u003cem\u003e33\u003c/em\u003e, 141–150, doi:10.1111/den.13688.\u003c/li\u003e\n\u003cli\u003eFukuda, H.; Ishihara, R.; Kato, Y.; Matsunaga, T.; Nishida, T.; Yamada, T.; Ogiyama, H.; Horie, M.; Kinoshita, K.; Tada, T. Comparison of performances of artificial intelligence versus expert endoscopists for real-time assisted diagnosis of esophageal squamous cell carcinoma (with video). \u003cem\u003eGastrointest Endosc\u003c/em\u003e \u003cstrong\u003e2020\u003c/strong\u003e, \u003cem\u003e92\u003c/em\u003e, 848–855, doi:10.1016/j.gie.2020.05.043.\u003c/li\u003e\n\u003cli\u003eMuto, M.; Minashi, K.; Yano, T.; Saito, Y.; Oda, I.; Nonaka, S.; Omori, T.; Sugiura, H.; Goda, K.; Kaise, M.; et al. Early detection of superficial squamous cell carcinoma in the head and neck region and esophagus by narrow band imaging: a multicenter randomized controlled trial. \u003cem\u003eJ Clin Oncol\u003c/em\u003e \u003cstrong\u003e2010\u003c/strong\u003e, \u003cem\u003e28\u003c/em\u003e, 1566–1572, doi:10.1200/JCO.2009.25.4680.\u003c/li\u003e\n\u003cli\u003eGrigoroiu, A.; Yoon, J.; Bohndiek, S.E. Deep learning applied to hyperspectral endoscopy for online spectral classification. \u003cem\u003eSci Rep\u003c/em\u003e \u003cstrong\u003e2020\u003c/strong\u003e, \u003cem\u003e10\u003c/em\u003e, 3947, doi:10.1038/s41598-020-60574-6.\u003c/li\u003e\n\u003cli\u003eMaktabi, M.; Wichmann, Y.; Kohler, H.; Ahle, H.; Lorenz, D.; Bange, M.; Braun, S.; Gockel, I.; Chalopin, C.; Thieme, R. Tumor cell identification and classification in esophageal adenocarcinoma specimens by hyperspectral imaging. \u003cem\u003eSci Rep\u003c/em\u003e \u003cstrong\u003e2022\u003c/strong\u003e, \u003cem\u003e12\u003c/em\u003e, 4508, doi:10.1038/s41598-022-07524-6.\u003c/li\u003e\n\u003cli\u003eKhan, M.J.; Khan, H.S.; Yousaf, A.; Khurshid, K.; Abbas, A. Modern Trends in Hyperspectral Image Analysis: A Review. \u003cem\u003eIEEE Access\u003c/em\u003e \u003cstrong\u003e2018\u003c/strong\u003e, \u003cem\u003e6\u003c/em\u003e, 14118–14129, doi:10.1109/access.2018.2812999.\u003c/li\u003e\n\u003cli\u003eCollins, T.; Maktabi, M.; Barberio, M.; Bencteux, V.; Jansen-Winkeln, B.; Chalopin, C.; Marescaux, J.; Hostettler, A.; Diana, M.; Gockel, I. Automatic Recognition of Colon and Esophagogastric Cancer with Machine Learning and Hyperspectral Imaging. \u003cem\u003eDiagnostics (Basel)\u003c/em\u003e \u003cstrong\u003e2021\u003c/strong\u003e, \u003cem\u003e11\u003c/em\u003e, doi:10.3390/diagnostics11101810.\u003c/li\u003e\n\u003cli\u003eMa, J.; Sun, D.W.; Pu, H.; Cheng, J.H.; Wei, Q. Advanced Techniques for Hyperspectral Imaging in the Food Industry: Principles and Recent Applications. \u003cem\u003eAnnu Rev Food Sci Technol\u003c/em\u003e \u003cstrong\u003e2019\u003c/strong\u003e, \u003cem\u003e10\u003c/em\u003e, 197–220, doi:10.1146/annurev-food-032818-121155.\u003c/li\u003e\n\u003cli\u003eYoon, J. Hyperspectral Imaging for Clinical Applications. \u003cem\u003eBioChip Journal\u003c/em\u003e \u003cstrong\u003e2022\u003c/strong\u003e, \u003cem\u003e16\u003c/em\u003e, 1–12, doi:10.1007/s13206-021-00041-0.\u003c/li\u003e\n\u003cli\u003evan Schaik, J.E.; Halmos, G.B.; Witjes, M.J.H.; Plaat, B.E.C. An overview of the current clinical status of optical imaging in head and neck cancer with a focus on Narrow Band imaging and fluorescence optical imaging. \u003cem\u003eOral Oncol\u003c/em\u003e \u003cstrong\u003e2021\u003c/strong\u003e, \u003cem\u003e121\u003c/em\u003e, 105504, doi:10.1016/j.oraloncology.2021.105504.\u003c/li\u003e\n\u003cli\u003eKurumi, H.; Nonaka, K.; Ikebuchi, Y.; Yoshida, A.; Kawaguchi, K.; Yashima, K.; Isomoto, H. Fundamentals, Diagnostic Capabilities and Perspective of Narrow Band Imaging for Early Gastric Cancer. \u003cem\u003eJ Clin Med\u003c/em\u003e \u003cstrong\u003e2021\u003c/strong\u003e, \u003cem\u003e10\u003c/em\u003e, doi:10.3390/jcm10132918.\u003c/li\u003e\n\u003cli\u003eYang, K.-Y.; Mukundan, A.; Tsao, Y.-M.; Shi, X.-H.; Huang, C.-W.; Wang, H.-C. Evaluating Hyperspectral Techniques Using Objective Metrics Research on Analog Narrowband Image. \u003cstrong\u003e2023\u003c/strong\u003e.\u003c/li\u003e\n\u003cli\u003eYao, K. Principles of magnifying endoscopy with narrow-band imaging. In \u003cem\u003eZoom Gastroscopy: Magnifying Endoscopy in the Stomach\u003c/em\u003e; Springer: 2013; pp. 49–56.\u003c/li\u003e\n\u003cli\u003eKawada, K.; Arima, M.; Miyahara, R.; Tsunomiya, M.; Kikuchi, M.; Yamamoto, F.; Hoshino, A.; Nakajima, Y.; Kinugasa, Y.; Kawano, T. Effect of adding magnifying BLI, magnifying NBI, and iodine staining to white light imaging in diagnosis of early esophageal cancer. \u003cem\u003eEndosc Int Open\u003c/em\u003e \u003cstrong\u003e2021\u003c/strong\u003e, \u003cem\u003e9\u003c/em\u003e, E1877-E1885, doi:10.1055/a-1583-9196.\u003c/li\u003e\n\u003cli\u003eTsai, T.J.; Mukundan, A.; Chi, Y.S.; Tsao, Y.M.; Wang, Y.K.; Chen, T.H.; Wu, I.C.; Huang, C.W.; Wang, H.C. Intelligent Identification of Early Esophageal Cancer by Band-Selective Hyperspectral Imaging. \u003cem\u003eCancers (Basel)\u003c/em\u003e \u003cstrong\u003e2022\u003c/strong\u003e, \u003cem\u003e14\u003c/em\u003e, doi:10.3390/cancers14174292.\u003c/li\u003e\n\u003cli\u003eTsai, C.L.; Mukundan, A.; Chung, C.S.; Chen, Y.H.; Wang, Y.K.; Chen, T.H.; Tseng, Y.S.; Huang, C.W.; Wu, I.C.; Wang, H.C. Hyperspectral Imaging Combined with Artificial Intelligence in the Early Detection of Esophageal Cancer. \u003cem\u003eCancers (Basel)\u003c/em\u003e \u003cstrong\u003e2021\u003c/strong\u003e, \u003cem\u003e13\u003c/em\u003e, doi:10.3390/cancers13184593.\u003c/li\u003e\n\u003cli\u003eSohan, M.; Sai Ram, T.; Rami Reddy, C.V. A review on yolov8 and its advancements. In Proceedings of the International Conference on Data Intelligence and Cognitive Informatics, 2024; pp. 529–545.\u003c/li\u003e\n\u003cli\u003eJiang, P.; Ergu, D.; Liu, F.; Cai, Y.; Ma, B. A Review of Yolo algorithm developments. \u003cem\u003eProcedia computer science\u003c/em\u003e \u003cstrong\u003e2022\u003c/strong\u003e, \u003cem\u003e199\u003c/em\u003e, 1066–1073.\u003c/li\u003e\n\u003cli\u003eTerven, J.; Córdova-Esparza, D.-M.; Romero-González, J.-A. A Comprehensive Review of YOLO Architectures in Computer Vision: From YOLOv1 to YOLOv8 and YOLO-NAS. \u003cem\u003eMachine Learning and Knowledge Extraction\u003c/em\u003e \u003cstrong\u003e2023\u003c/strong\u003e, \u003cem\u003e5\u003c/em\u003e, 1680–1716, doi:10.3390/make5040083.\u003c/li\u003e\n\u003cli\u003ePham, T.-L.; Le, V.-H. Ovarian Tumors Detection and Classification from Ultrasound Images Based on YOLOv8. \u003cem\u003eJournal of Advances in Information Technology\u003c/em\u003e \u003cstrong\u003e2024\u003c/strong\u003e, \u003cem\u003e15\u003c/em\u003e, 264–275, doi:10.12720/jait.15.2.264-275.\u003c/li\u003e\n\u003cli\u003eSolimani, F.; Cardellicchio, A.; Dimauro, G.; Petrozza, A.; Summerer, S.; Cellini, F.; Renò, V. Optimizing tomato plant phenotyping detection: Boosting YOLOv8 architecture to tackle data complexity. \u003cem\u003eComputers and Electronics in Agriculture\u003c/em\u003e \u003cstrong\u003e2024\u003c/strong\u003e, \u003cem\u003e218\u003c/em\u003e, doi:10.1016/j.compag.2024.108728.\u003c/li\u003e\n\u003cli\u003eLakshmi, A.J.; Kishore, G.; Kumar, T.D.; Koushik, V. Drone detection and classification using YOLOv8 and deep CNN. In Proceedings of the International Conference on Innovations and Advances in Cognitive Systems, 2024; pp. 14–27.\u003c/li\u003e\n\u003cli\u003eAhmed, M.; Afreen, N.; Ahmed, M.; Sameer, M.; Ahamed, J. An inception V3 approach for malware classification using machine learning and transfer learning. \u003cem\u003eInternational Journal of Intelligent Networks\u003c/em\u003e \u003cstrong\u003e2023\u003c/strong\u003e, \u003cem\u003e4\u003c/em\u003e, 11–18, doi:10.1016/j.ijin.2022.11.005.\u003c/li\u003e\n\u003cli\u003eKim, J.-H. Improvement of Inceptionv3 Model Classification Performance Using Chest X-Ray Images. \u003cem\u003eJournal of Mechanics in Medicine and Biology\u003c/em\u003e \u003cstrong\u003e2022\u003c/strong\u003e, \u003cem\u003e22\u003c/em\u003e, doi:10.1142/s0219519422400322.\u003c/li\u003e\n\u003cli\u003eMujahid, M.; Rustam, F.; Alvarez, R.; Luis Vidal Mazon, J.; Diez, I.T.; Ashraf, I. Pneumonia Classification from X-ray Images with Inception-V3 and Convolutional Neural Network. \u003cem\u003eDiagnostics (Basel)\u003c/em\u003e \u003cstrong\u003e2022\u003c/strong\u003e, \u003cem\u003e12\u003c/em\u003e, doi:10.3390/diagnostics12051280.\u003c/li\u003e\n\u003cli\u003eXia, X.; Xu, C.; Nan, B. Inception-v3 for flower classification. In Proceedings of the 2017 2nd international conference on image, vision and computing (ICIVC), 2017; pp. 783–787.\u003c/li\u003e\n\u003cli\u003eLiu, K.; Yu, S.; Liu, S. An Improved InceptionV3 Network for Obscured Ship Classification in Remote Sensing Images. \u003cem\u003eIEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing\u003c/em\u003e \u003cstrong\u003e2020\u003c/strong\u003e, \u003cem\u003e13\u003c/em\u003e, 4738–4747, doi:10.1109/jstars.2020.3017676.\u003c/li\u003e\n\u003cli\u003eMuhathir; Dwi Ryandra, M.F.; Syah, R.B.; Khairina, N.; Muliono, R. Convolutional neural network (CNN) of Resnet-50 with Inceptionv3 architecture in classification on X-ray image. In Proceedings of the Computer Science On-line Conference, 2023; pp. 208–221.\u003c/li\u003e\n\u003cli\u003eZ. Riaz, B. Khan, S. Abdullah, S. Khan, and M. S. Islam, “Lung Tumor Image Segmentation from Computer Tomography Images Using MobileNetV2 and Transfer Learning,” \u003cem\u003eBioengineering\u003c/em\u003e, vol. 10, no. 8, p. 981, Aug. 2023, doi: 10.3390/bioengineering10080981.\u003c/li\u003e\n\u003cli\u003eM. S. Islam Sajol, A. S. M Jahid Hasan, M. S. Islam, and M. S. Rahman, “A ConvNeXt V2 Approach to Document Image Analysis: Enhancing High-Accuracy Classification,” in \u003cem\u003e2024 IEEE 3rd Conference on Information Technology and Data Science (CITDS)\u003c/em\u003e, Debrecen, Hungary: IEEE, Aug. 2024, pp. 1–6. doi: 10.1109/CITDS62610.2024.10791343.\u003c/li\u003e\n\u003cli\u003eNguyen, H. Fast object detection framework based on mobilenetv2 architecture and enhanced feature pyramid. \u003cem\u003eJ. Theor. Appl. Inf. Technol\u003c/em\u003e \u003cstrong\u003e2020\u003c/strong\u003e, \u003cem\u003e98\u003c/em\u003e, 812–824.\u003c/li\u003e\n\u003cli\u003eDong, K.; Zhou, C.; Ruan, Y.; Li, Y. MobileNetV2 Model for Image Classification. In Proceedings of the 2020 2nd International Conference on Information Technology and Computer Application (ITCA), 2020; pp. 476–480.\u003c/li\u003e\n\u003cli\u003eHamid, Y.; Wani, S.; Soomro, A.B.; Alwan, A.A.; Gulzar, Y. Smart Seed Classification System based on MobileNetV2 Architecture. In Proceedings of the 2022 2nd International Conference on Computing and Information Technology (ICCIT), 2022; pp. 217–222.\u003c/li\u003e\n\u003cli\u003eKhasoggi, B.; Ermatita, E.; Samsuryadi, S. Efficient mobilenet architecture as image recognition on mobile and embedded devices. \u003cem\u003eIndonesian Journal of Electrical Engineering and Computer Science\u003c/em\u003e \u003cstrong\u003e2019\u003c/strong\u003e, \u003cem\u003e16\u003c/em\u003e, doi:10.11591/ijeecs.v16.i1.pp389-394.\u003c/li\u003e\n\u003cli\u003eGulzar, Y. Fruit Image Classification Model Based on MobileNetV2 with Deep Transfer Learning Technique. \u003cem\u003eSustainability\u003c/em\u003e \u003cstrong\u003e2023\u003c/strong\u003e, \u003cem\u003e15\u003c/em\u003e, doi:10.3390/su15031906.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Deep Learning, Medical Image Processing, Esophageal Cancer, Spectrum-Aided Vision Enhancer (SAVE), Virtual NBI, Computer-Aided Diagnosis (CAD), Object Detection","lastPublishedDoi":"10.21203/rs.3.rs-8677211/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8677211/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eEsophageal cancer is a highly aggressive malignancy where early detection is critical for survival. However, early-stage lesions typically present subtle mucosal changes that are difficult to identify using standard White Light Imaging (WLI), and hardware-based Narrow Band Imaging (NBI) is not universally available. In this study, we propose a novel image processing algorithm termed the Spectrum-Aided Vision Enhancer (SAVE) to address these challenges in computer-aided diagnosis (CAD). Leveraging hyperspectral data principles, SAVE transforms standard WLI endoscopic images into enhanced, NBI-like representations, significantly improving mucosal contrast and lesion visibility without requiring additional hardware. To validate the efficacy of this approach for medical image analysis, we utilized a dataset of Squamous Cell Carcinoma (SCC) and dysplasia. We conducted a comprehensive comparative analysis using five state-of-the-art deep learning models: YOLOv8, InceptionV3, Inception-ResNet-V2, ConvNeXt-V2, and MobileNetV2. Experimental results demonstrate that models trained on SAVE-enhanced images significantly outperform those trained on traditional WLI in both classification and detection tasks. This study presents a cost-effective, software-driven solution that integrates advanced image processing with deep learning, offering a robust tool for the automated screening of esophageal malignancies.\u003c/p\u003e","manuscriptTitle":"Enhancing Esophageal Cancer Detection via Virtual NBI: A Novel Spectrum-Aided Vision Enhancer (SAVE) and Deep Learning Framework","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-23 17:33:31","doi":"10.21203/rs.3.rs-8677211/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-03-03T05:51:05+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-02T17:16:40+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-01T10:11:07+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"261798591302151652062671519735853815443","date":"2026-02-25T12:19:24+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"314387908272011738813485002934344427636","date":"2026-02-17T06:17:44+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-02-16T22:42:42+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-02-16T22:38:17+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-02-13T05:50:41+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-01-28T13:37:01+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2026-01-28T12:16:19+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"ee44baa3-80c7-4b7c-894f-2084f0550916","owner":[],"postedDate":"February 23rd, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":63047684,"name":"Biological sciences/Cancer"},{"id":63047685,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":63047686,"name":"Physical sciences/Engineering"},{"id":63047687,"name":"Physical sciences/Mathematics and computing"}],"tags":[],"updatedAt":"2026-05-06T05:24:33+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-23 17:33:31","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8677211","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8677211","identity":"rs-8677211","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}