GoogLeNet/DenseNet-201 to classify near-infrared (NIR) spectrum graphs for cancer diagnosis – using pretrained image networks for medical spectroscopy

GoogLeNet/DenseNet-201 to classify near-infrared (NIR) spectrum graphs for cancer diagnosis – using pretrained image networks for medical spectroscopy | 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 GoogLeNet/DenseNet-201 to classify near-infrared (NIR) spectrum graphs for cancer diagnosis – using pretrained image networks for medical spectroscopy Tanmoy Bhattacharjee This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6562812/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract The study compares sensitivity/specificity of classification by pretrained image networks and traditional Machine Learning (ML) methods. One hundred seven spectra each of benign skin conditions actinic keratosis (ACK) and seborrheic keratosis (SEK), and skin cancer basal cell carcinoma (BCC) were downloaded from a public database. Eighty spectra per group were used for training and twenty-seven spectra per group for testing. In the first classification strategy, spectrum intensity values were used as input for Linear Discriminant Analysis (LDA), K-Nearest Neighbors (KNN), Decision Tree (DT), TreeBagger, Ensemble method, Naïve Bayes, Support Vector Machine (SVM), and Artificial Neural Network (ANN). The second strategy involved using spectrum graphs saved as images to train GoogLeNet, Places-365 GoogLeNet, ResNet-50, Inception-V3, DenseNet-201, and NasNetMobile. Strategy 2 yielded better sensitivity/specificity – 0.7/ 0.91 (ACK), 0.7/0.83 (BCC), and 0.63/0.85 (SEK) compared to strategy 1–0.52/0.94 (ACK), 0.7/0.8 (BCC), and 0.5/0.8 (SEK). Grad-CAM mapping suggested that 1100–1200,1350–1450, and 1600–1700 1/cm to be responsible for classification by strategy 2. When these regions plotted as subplots and saved as images were used for training using strategy 2, sensitivity for BCC increases to 0.78. Results suggest using pretrained image networks to classify spectra may yield better results, give a visual understanding of the basis of classification, and provide means to improve classification further. Physical sciences/Optics and photonics/Other photonics/Biophotonics Physical sciences/Optics and photonics/Optical techniques/Optical spectroscopy/Near-infrared spectroscopy Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Near-infrared spectroscopy (NIR), attenuated total reflection Fourier-transform infrared spectroscopy (ATR-FTIR), Raman spectroscopy, and Fluorescence spectroscopy are biospectroscopy systems that has enormous potential for medical diagnosis in point-of-care settings 1 – 5 . Reagent-free, requiring minimal sample preparation, often amenable for in-vivo applications, these rapid techniques are especially suited for routine clinical practices. Hence, they have been explored extensively for diagnosis of cancer, neurodegenerative diseases, cardiovascular complications, developmental disorders, pathogen identification, and many other areas. The output of measurement by a biospectroscopy system is a spectrum. For medical applications, the spectra obtained from sample (organ, tissue, cells, biofluid) are fed into chemometric algorithms 6 . For example, to differentiate normal tissue from cancer, spectra are obtained from normal tissue and cancer tissue, and all these spectra are inputted into an algorithm for pattern recognition and classification. There are several pattern recognition and machine learning tools that are available - Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), K-Nearest Neighbors (KNN), Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Network (ANN) to name a few. However, these do not always give the best classification results, especially when many groups need to be classified. For example, ideal result for normal vs cancer will be classification of all normal spectra from cancer spectra (100% sensitivity). However, 100% sensitivity is rarely obtained. When number of groups increase, for example when classifying normal, benign condition, inflammatory lesion, and cancer spectra, the sensitivity may be less than 60%. This is far from ideal for medical diagnosis. Strategies used to mitigate the problem are increasing the sample size and optimizing the hyperparameters. However, both might fail to achieve the required results. A new analysis strategy is thus needed. Using pre-trained image convolutional neural network (CNN) architecture is one such strategy. CNN is a popular deep learning algorithm, wherein the model learns patterns directly from images and uses them to classify images. Pre-trained CNNs are CNNs that have been trained using millions of images to classify up to 1000 groups/categories of image. These can be adapted to specific classification tasks (example cancer image vs normal image) by method called transfer learning 7 , 8 . Pre-trained CNNs are superior to traditional machine learning techniques, owing to their self-learning and autonomous feature extraction capabilities. However, since the input for pre-trained CNNs are images, these have not been extensively explored for spectra classification. In the current study, we use pretrained image CNNs for classification. Each spectrum is plotted, saved as an image, and used as input for GoogLeNet, Places-365 GoogLeNet, ResNet-50, Inception-V3, DenseNet-201, and NasNetMobile. The method is used to classify three skin cancer sub-types spectra available in a public database. Identification of features responsible for classification (Explainable AI) 9 and using the same to improve classification is discussed. Methods Data : The near-infrared (NIR) spectra available in the NIR-SC-UFES Dataset 10 was used for the study. The dataset has non-cancers − 336 ACK (actinic keratosis), 188 SEK (seborrheic keratosis), and 62 NEV (nevus) spectra, and cancers − 302 BCC (basal cellular carcinoma), 72 SCC (squamous cellular carcinoma), and 11 MEL (melanoma) spectra. For the current study, 80 BCC, ACK, and SEK spectra each were used for training and 27 spectra each were used for testing. Classification using traditional machine learning methods (strategy 1) : The spectra were normalized using standard normal variate (SNV) and used as input for Linear Discriminant Analysis (LDA), K-Nearest Neighbors (KNN), Decision Tree (DT), TreeBagger, Ensemble method, Naïve Bayes, Support Vector Machine (SVM), and Artificial Neural Network (ANN). LDA and SVM models were also optimized by optimizing the hyperparameters. All models were tested using the test data. Classification using pre-trained image network (strategy 2) : Each spectrum was normalized using standard normal variate (SNV) method, plotted along with their axes and saved as images (‘.png’ format). The spectra were then put into respective folders and sub-folders – Training – BCC, ACK, and SEK and Test - – BCC, ACK, and SEK. Image datastores were created using these folders and subfolders. The pre-trained CNNs were prepared for transfer learning by replacing their classification layer and changing weight learn rate factor and bias learn rate factor to 10 in the learnable layers. The image datastores were augmented by resizing them based on CNN input size and color processing (gray to RGB) before using them as input for CNNs. Six CNNs, GoogLeNet, Places-365 GoogLeNet, ResNet-50, Inception-V3, DenseNet-201, and NasNetMobile, were trained using the augmented training image store and tested using the test datastore. Training was performed using three mini batch sizes – 8, 10, and 12. The initial learn rate was 0.00001. Solver type used was Stochastic Gradient Descent with Momentum (sgdm). After finding the best CNN and mini batch size for classification, the CNN was trained multiple times (20 times). Gradient-weighted class activation mapping (Grad-CAM) 9 was used to find regions of the images used for classification by the CNN. Spectral ranges identified in this manner were plotted as subplots with (strategy 3) or without (strategy 4) axes, saved as images, and fed into CNNs to study improvement in classification, if any. Classification parameters : Precision, sensitivity, specificity, accuracy, and F-scores were calculated using MATLAB script published by Bertolini 11 . All analyses were performed using MATLAB 2024b. Results Classification of ACK, BCC, and SEK with spectra intensity values as input using SVM gives the best precision (0.82) and recall (0.94) for ACK, but sensitivity is low (0.52). NB gives lower precision (0.56) and recall (0.74) for ACK, but sensitivity is slightly higher (0.67). SVM with hyperparameters optimized gives best precision (0.63), sensitivity (0.7), and recall (0.8) for BCC. SVM and SVM with hyperparameters optimized give best results (precision/sensitivity/recall ~ 0.6/0.5/0.8) for SEK (Table 1 , strategy 1). By using images of spectra (Fig. 1 ) as input (Table 1 , strategy 2) for pre-trained network DenseNet-201 (round 18, minibatch size 10), precision, sensitivity, and recall were ~ 0.8, 0.7, and 0.91, respectively for ACK. GoogLeNet (round 18, minibatch size 8) gives the best results for BCC (precision/sensitivity/recall – 0.68/0.7/0.83). DenseNet-201 (round 4, minibatch size 10) gives best precision/sensitivity/recall (0.68/0.63/0.85) for SEK. Since the networks were trained 20 times independently, the training round wherein the above results were obtained is mentioned. For example, DenseNet-201 (round 4, minibatch size 10) means that when the DenseNet-201 network was trained for the 4th time (independently) using the same training and test data, precision/sensitivity/recall of 0.68/0.63/0.85 was obtained for SEK. To understand features that contributed to the classification, Grad-CAM was applied to all test images (Fig. 2 ). The 1100–1200 1/cm and 1350–1450 1/cm regions were responsible for 10 out of 19 and 4 out of 19 ACK spectra image classification (Fig. 2 a). In 5 images, sections in the background seems to have been responsible for ACK classification. 10 out of 19 BCC image classification were due to 1600–1700 1/cm region, 14 of 19 were due to 1100–1200 1/cm region, and none were due to background sections (Fig. 2 b). 12 out of 17 SEK images were classified due to 1600–1700 1/cm region while 4 of 17 images were classified due to 1100–1200 1/cm. When these spectral regions were plotted as subplots, saved as images (Fig. 3 ) and used as input for DenseNet-201 (mini batch size 8, round 5), improvement in sensitivity (0.78) for BCC was observed (Table 1 , strategy 3) compared to using entire range images as input (specificity – 0.7) (Table 1 , strategy 4). When the same spectral ranges plotted as subplots without axes (Fig. 4 ), saved as images, and used as input for GoogLeNet (round 10, mini batch size 10), improvement in precision (0.76) and specificity (0.89) was observed (Table 1 , method d)) compared to method b (precision/specificity – 0.68/0.83). Note that precision and specificity of ACK also increased to 1, but sensitivity was low (0.41). Table 1 Classification using different strategies 1) using spectra intensity values as input for traditional machine learning (ML) methods, 2. using spectra graphs saved as images as input for pretrained image networks, 3. using spectral regions identified by Grad-CAM mapping plotted as subplots with axes and saved as images as input for pretrained image networks, and 4. using spectral regions identified by Grad-CAM mapping plotted as subplots without axes and saved as images as input for pretrained image networks. SVM: Support Vector Machine. NB: Naïve Bayes. Opt SVM: SVM model with hyperparameters optimized. Method ML/DL Precision Sensitivity Specificity/Recall F-measure Accuracy Range (1/cm) ACK BCC SEK ACK BCC SEK ACK BCC SEK ACK BCC SEK ACK BCC SEK 1 SVM 0.82 0.54 0.59 0.52 0.70 0.63 0.94 0.70 0.78 0.64 0.61 0.61 0.62 0.62 0.62 900–1800 NB 0.56 0.53 0.56 0.67 0.78 0.19 0.74 0.65 0.93 0.61 0.63 0.28 0.54 0.54 0.54 Opt SVM 0.57 0.63 0.61 0.59 0.70 0.52 0.78 0.80 0.83 0.58 0.67 0.56 0.60 0.60 0.60 SVM 0.82 0.54 0.59 0.52 0.70 0.63 0.94 0.70 0.78 0.64 0.61 0.61 0.62 0.62 0.62 Opt SVM 0.57 0.63 0.61 0.59 0.70 0.52 0.78 0.80 0.83 0.58 0.67 0.56 0.60 0.60 0.60 2 DenseNet-201 0.79 0.63 0.52 0.70 0.56 0.63 0.91 0.83 0.70 0.75 0.59 0.57 0.63 0.63 0.63 GoogLeNet 0.50 0.68 0.59 0.48 0.70 0.59 0.76 0.83 0.80 0.49 0.69 0.59 0.59 0.59 0.59 DenseNet-201 0.68 0.61 0.68 0.63 0.70 0.63 0.85 0.78 0.85 0.65 0.66 0.65 0.65 0.65 0.65 3 DenseNet-201 0.59 0.66 0.5 0.63 0.78 0.37 0.78 0.8 0.81 0.61 0.71 0.43 0.59 0.59 0.59 1100–1200, 1350–1450,1600–1700 4 GoogLeNet 1 0.76 0.56 0.41 0.70 0.93 1 0.89 0.63 0.58 0.73 0.69 0.68 0.68 0.68 1100–1200, 1350–1450,1600–1700 Discussion A search of keywords “spectroscopy medical diagnosis classification” in Google Scholar revealed roughly 180 articles since 2025 in the first 20 pages itself, pertaining to use of biospectroscopy for medical disease diagnosis using machine learning (ML)-based classification. From experience of working with biospectroscopy groups, many more articles are not submitted or fail to get published because they do not achieve good classification results. Especially with multi-class classification problems, where number of groups exceed six or ten, getting good classification for all classes get tougher. A new method of classifying spectra is hence required. Pretrained image networks are trained to classify images belong to thousand or more different groups and are well-suited to tackle multi-class classification problems with large number of groups. Woo et.al. have used images of graphs of multi-sensor signals for classification of sleep stages 12 . Williams et. al. used of GoogLeNet and ResNet for classification of microplastics Raman and FTIR spectra using images of spectrum graphs 13 . In the current study, we explored three novel aspects of using spectral graph images as input for pretrained image networks – 1. whether this strategy is better than traditional ML methods such as SVM, LDA, and so on, 2. whether this strategy can give us the features used by networks for classification, and 3. whether these features can be selectively used to improve classification. We found that pretrained image networks give better classification than traditional ML, and that spectral regions responsible for classification can be identified and used to improve classification. The study was done using a small dataset, hence there can be a question on whether the results will be true for large datasets. However, most biomedical researchers start with small datasets to explore study feasibility, and poor classification tend to discourage researchers from moving to larger studies. Hence, testing on small dataset may be more apt. It should be noted that one method did not give good classification for all groups. SVM without any optimization gave the best classification for ACK, while SVM post optimization gave best results for BCC and SEK. Similarly, DenseNet-201 gave the best results for ACK and SEK, while GoogLeNet gave the best results for BCC. We envision using different trained models in parallel for classification in real-world applications. For example, to identify whether an unknown sample is ACK, BCC, or SEK, both DenseNet-201 and GoogLeNet will be used for prediction. If the predictions do not match, weightage will be given to the network best trained for predicted group. For example, if DenseNet predicts ACK, and GoogLeNet predicts SEK, the verdict will be ACK as DenseNet predicts ACK better. Regarding feature identification, pretrained image networks have an edge over traditional ML. PCA loading factors or Shapley values used to identify features for ML methods give underlying features in general, but not for each spectrum. With pretrained networks, Grad-CAM mapping allows feature identification in each spectrum image (Fig. 2 ). In Fig. 2 a, some ACK images are classified due to 1100–1200 1/cm region while others were due to 1350–1450 1/cm. This allows better understanding of the basis of classification. Further, it is observed that some images have been classified based on background. Their predictions can be marked as “suspect”. During decision making, as in the example in the previous paragraph, if DenseNet predicts ACK, but basis of classification is image background, while GoogLeNet predicts SEK, where the basis of classification is the spectral region, then verdict will be SEK. Once spectral regions are identified, graphs of regions plotted as subplots could be used to improve sensitivity/specificity. There are multiple ways to do this, one of which was explored in the study – subplots without axes. Other options are to arrange subplots in horizontal manner instead of vertically. Each region may be further magnified by sub plotting sections. For example, instead of subplots of 1100–1200 and 1350–1450 1/cm, subplots of 1100–1150, 1150–1200, 1350–1400, and 1400–1450 1/cm can be generated and used as input. MATLAB allows “no spacing option” in “tiledlayout” to remove space between subplots. Simple algorithms can be generated to further remove white space, if any. Another advantage of pretrained networks is that they can be trained multiple times. The results of training vary. This allows the best trained model to be chosen; the best model being the one that gives the highest precision/sensitivity/specificity and where most or all images have been classified based on spectral region instead of background. In this study, each network was trained 20 times to choose the best trained model. Further, training options can be varied. Stochastic Gradient Descent with Momentum (sgdm), Root Mean Square Propagation, or Adaptive Moment Estimation, etc. can be used as solvers. In this study, only ‘sgdm’ was used, but other methods can be explored to improve classification. The mini batch size can be varied. In the current study, bath sizes 8, 10, and 12 were used, each giving different results. While the initial learn rate was kept 0.00001 in this study, this can be varied to improve results. Finally, the process can be fully automated. We developed a MATLAB script to automatically import and preprocess spectra, plot spectrum/spectrum ranges/subplots as per user input, save these as images in automatically generated labelled subfolders, import images with label information into image datastores, train and test different pretrained networks, save trained models, and collate classification outcomes in an excel sheet. The code can be made available on request. References Martin FL (2023) Translating Biospectroscopy Techniques to Clinical Settings: A New Paradigm in Point-of-Care Screening and/or Diagnostics. J Personalised Med 13:1511 Kannan S, Callery EL, Rowbottom AW (2023) Vibrational biospectroscopy in the clinical setting: Exploring the impact of new advances in the field of immunology. Journal of Spectroscopy. 5557441 (2023) Kim Y, Bui TT, Chung H (2025) A review on diverse spectroscopic methods for identification of hepato-pancreato-biliary diseases using human bile as a specimen. Appl Spectrosc Rev, 1–20 Son Y et al (2025) Advances in spectroscopic detection of traumatic brain injury biomarkers: Potential for diagnostic applications. Appl Spectrosc Rev, 1–30 Delrue C, De Bruyne S, Speeckaert MM (2025) The Promise of Infrared Spectroscopy in Liquid Biopsies for Solid Cancer Detection. Diagnostics 15:368 Xuesong H et al (2024) Commentary on the review articles of spectroscopy technology combined with chemometrics in the last three years. Appl Spectrosc Rev 59:423–482 Spolaôr N et al (2024) Fine-tuning pre-trained neural networks for medical image classification in small clinical datasets. Multimedia Tools Appl 83:27305–27329 Archana R, Jeevaraj PE (2024) Deep learning models for digital image processing: a review. Artif Intell Rev 57:11 Raghavan K (2024) Attention guided grad-CAM: an improved explainable artificial intelligence model for infrared breast cancer detection. Multimedia Tools Appl 83:57551–57578 da Cunha PHP et al (2027) NIR-SC-UFES: A portable NIR spectral dataset to skin cancer in vivo. medRxiv, 2024.2011. 24317165 (2024) Bertolini E, Precision, Specificity, Sensitivity A (2025) & F1-score ( https://www.mathworks.com/matlabcentral/fileexchange/86158-precision-specificity-sensitivity-accuracy-f1-score) . MATLAB Central File Exchange. Retrieved April 24, (2025) Woo Y et al (2023) Automatic Sleep Stage Classification Using Deep Learning Algorithm for Multi-Institutional Database. IEEE Access 11:46297–46307 Williams WA, Arunprasad A, Aravamudhan S (2024) Application of a modified set of GoogLeNet and ResNet-18 convolutional neural networks towards the identification of environmentally derived-MPLs in the Yadkin-pee dee river basin. Environ Syst Res 13:53 Additional Declarations There is NO Competing Interest. Cite Share Download PDF Status: Posted Version 1 posted 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-6562812","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":451252011,"identity":"4023de05-de44-40bc-8932-ed53ddbfc200","order_by":0,"name":"Tanmoy Bhattacharjee","email":"data:image/png;base64,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","orcid":"https://orcid.org/0000-0002-2550-3136","institution":"Oudari Consultancy","correspondingAuthor":true,"prefix":"","firstName":"Tanmoy","middleName":"","lastName":"Bhattacharjee","suffix":""}],"badges":[],"createdAt":"2025-04-30 08:31:37","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6562812/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6562812/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":82070451,"identity":"1ccebd6b-9e20-4a17-88a7-bfc96456affc","added_by":"auto","created_at":"2025-05-06 13:11:29","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":27590,"visible":true,"origin":"","legend":"\u003cp\u003eNIR Spectra of ACK, BCC, and SEK from the database\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6562812/v1/be61905a585aef4ca4b4f31a.png"},{"id":82070957,"identity":"13873a57-d9ee-4764-a753-94511da437ac","added_by":"auto","created_at":"2025-05-06 13:19:29","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1742871,"visible":true,"origin":"","legend":"\u003cp\u003eGrad-CAM mapping of test images a) ACK (DenseNet-201 trained network), b) BCC (GoogLeNet trained network), and c) SEK (DenseNet-201 trained network). Note: The blanks are images where the predictions by trained network were incorrect. They were omitted to ensure only spectral regions of correctly classified images were used for next steps.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-6562812/v1/fa54bdfa57d8147d4b405806.png"},{"id":82070956,"identity":"abeeb7c2-30de-4ef9-91ff-f2fb9116dd37","added_by":"auto","created_at":"2025-05-06 13:19:29","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":96387,"visible":true,"origin":"","legend":"\u003cp\u003eSpectral regions plotted as subplots with axes and saved as images for input into pretrained networks a) ACK, b) BCC, and c) SEK\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-6562812/v1/1195fc3708bbfcda4274e4a9.png"},{"id":82070453,"identity":"b792cd73-9bb1-479d-ba10-fd13ac79d5d1","added_by":"auto","created_at":"2025-05-06 13:11:29","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":21413,"visible":true,"origin":"","legend":"\u003cp\u003eSpectral regions plotted as subplots without axes and saved as images for input into pretrained networks a) ACK, b) BCC, and c) SEK\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-6562812/v1/dcc2da0ee891524d5dba31e0.png"},{"id":82070993,"identity":"95adc9e2-7dc4-4c17-9ebe-90ba6aaeb44a","added_by":"auto","created_at":"2025-05-06 13:19:35","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2903289,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6562812/v1/67c1a4b0-ef55-4cb3-bd12-310adabbc578.pdf"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e Competing Interest.","formattedTitle":"GoogLeNet/DenseNet-201 to classify near-infrared (NIR) spectrum graphs for cancer diagnosis – using pretrained image networks for medical spectroscopy","fulltext":[{"header":"Introduction","content":"\u003cp\u003eNear-infrared spectroscopy (NIR), attenuated total reflection Fourier-transform infrared spectroscopy (ATR-FTIR), Raman spectroscopy, and Fluorescence spectroscopy are biospectroscopy systems that has enormous potential for medical diagnosis in point-of-care settings\u003csup\u003e\u003cspan additionalcitationids=\"CR2 CR3 CR4\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. Reagent-free, requiring minimal sample preparation, often amenable for in-vivo applications, these rapid techniques are especially suited for routine clinical practices. Hence, they have been explored extensively for diagnosis of cancer, neurodegenerative diseases, cardiovascular complications, developmental disorders, pathogen identification, and many other areas.\u003c/p\u003e \u003cp\u003eThe output of measurement by a biospectroscopy system is a spectrum. For medical applications, the spectra obtained from sample (organ, tissue, cells, biofluid) are fed into chemometric algorithms\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. For example, to differentiate normal tissue from cancer, spectra are obtained from normal tissue and cancer tissue, and all these spectra are inputted into an algorithm for pattern recognition and classification. There are several pattern recognition and machine learning tools that are available - Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), K-Nearest Neighbors (KNN), Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Network (ANN) to name a few. However, these do not always give the best classification results, especially when many groups need to be classified. For example, ideal result for normal vs cancer will be classification of all normal spectra from cancer spectra (100% sensitivity). However, 100% sensitivity is rarely obtained. When number of groups increase, for example when classifying normal, benign condition, inflammatory lesion, and cancer spectra, the sensitivity may be less than 60%. This is far from ideal for medical diagnosis. Strategies used to mitigate the problem are increasing the sample size and optimizing the hyperparameters. However, both might fail to achieve the required results. A new analysis strategy is thus needed.\u003c/p\u003e \u003cp\u003eUsing pre-trained image convolutional neural network (CNN) architecture is one such strategy. CNN is a popular deep learning algorithm, wherein the model learns patterns directly from images and uses them to classify images. Pre-trained CNNs are CNNs that have been trained using millions of images to classify up to 1000 groups/categories of image. These can be adapted to specific classification tasks (example cancer image vs normal image) by method called transfer learning\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e,\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e. Pre-trained CNNs are superior to traditional machine learning techniques, owing to their self-learning and autonomous feature extraction capabilities. However, since the input for pre-trained CNNs are images, these have not been extensively explored for spectra classification. In the current study, we use pretrained image CNNs for classification. Each spectrum is plotted, saved as an image, and used as input for GoogLeNet, Places-365 GoogLeNet, ResNet-50, Inception-V3, DenseNet-201, and NasNetMobile. The method is used to classify three skin cancer sub-types spectra available in a public database. Identification of features responsible for classification (Explainable AI)\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e and using the same to improve classification is discussed.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e \u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003eData\u003c/span\u003e: The near-infrared (NIR) spectra available in the NIR-SC-UFES Dataset\u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e was used for the study. The dataset has non-cancers \u0026minus;\u0026thinsp;336 ACK (actinic keratosis), 188 SEK (seborrheic keratosis), and 62 NEV (nevus) spectra, and cancers \u0026minus;\u0026thinsp;302 BCC (basal cellular carcinoma), 72 SCC (squamous cellular carcinoma), and 11 MEL (melanoma) spectra. For the current study, 80 BCC, ACK, and SEK spectra each were used for training and 27 spectra each were used for testing.\u003c/p\u003e \u003cp\u003e \u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003eClassification using traditional machine learning methods (strategy 1)\u003c/span\u003e: The spectra were normalized using standard normal variate (SNV) and used as input for Linear Discriminant Analysis (LDA), K-Nearest Neighbors (KNN), Decision Tree (DT), TreeBagger, Ensemble method, Na\u0026iuml;ve Bayes, Support Vector Machine (SVM), and Artificial Neural Network (ANN). LDA and SVM models were also optimized by optimizing the hyperparameters. All models were tested using the test data.\u003c/p\u003e \u003cp\u003e \u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003eClassification using pre-trained image network (strategy 2)\u003c/span\u003e: Each spectrum was normalized using standard normal variate (SNV) method, plotted along with their axes and saved as images (\u0026lsquo;.png\u0026rsquo; format). The spectra were then put into respective folders and sub-folders \u0026ndash; Training \u0026ndash; BCC, ACK, and SEK and Test - \u0026ndash; BCC, ACK, and SEK. Image datastores were created using these folders and subfolders. The pre-trained CNNs were prepared for transfer learning by replacing their classification layer and changing weight learn rate factor and bias learn rate factor to 10 in the learnable layers. The image datastores were augmented by resizing them based on CNN input size and color processing (gray to RGB) before using them as input for CNNs. Six CNNs, GoogLeNet, Places-365 GoogLeNet, ResNet-50, Inception-V3, DenseNet-201, and NasNetMobile, were trained using the augmented training image store and tested using the test datastore. Training was performed using three mini batch sizes \u0026ndash; 8, 10, and 12. The initial learn rate was 0.00001. Solver type used was Stochastic Gradient Descent with Momentum (sgdm). After finding the best CNN and mini batch size for classification, the CNN was trained multiple times (20 times). Gradient-weighted class activation mapping (Grad-CAM)\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e was used to find regions of the images used for classification by the CNN. Spectral ranges identified in this manner were plotted as subplots with (strategy 3) or without (strategy 4) axes, saved as images, and fed into CNNs to study improvement in classification, if any.\u003c/p\u003e \u003cp\u003e \u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003eClassification parameters\u003c/span\u003e: Precision, sensitivity, specificity, accuracy, and F-scores were calculated using MATLAB script published by Bertolini\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eAll analyses were performed using MATLAB 2024b.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eClassification of ACK, BCC, and SEK with spectra intensity values as input using SVM gives the best precision (0.82) and recall (0.94) for ACK, but sensitivity is low (0.52). NB gives lower precision (0.56) and recall (0.74) for ACK, but sensitivity is slightly higher (0.67). SVM with hyperparameters optimized gives best precision (0.63), sensitivity (0.7), and recall (0.8) for BCC. SVM and SVM with hyperparameters optimized give best results (precision/sensitivity/recall\u0026thinsp;~\u0026thinsp;0.6/0.5/0.8) for SEK (Table \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e, strategy 1).\u003c/p\u003e\n\u003cp\u003eBy using images of spectra (Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e) as input (Table \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e, strategy 2) for pre-trained network DenseNet-201 (round 18, minibatch size 10), precision, sensitivity, and recall were ~\u0026thinsp;0.8, 0.7, and 0.91, respectively for ACK. GoogLeNet (round 18, minibatch size 8) gives the best results for BCC (precision/sensitivity/recall \u0026ndash; 0.68/0.7/0.83). DenseNet-201 (round 4, minibatch size 10) gives best precision/sensitivity/recall (0.68/0.63/0.85) for SEK. Since the networks were trained 20 times independently, the training round wherein the above results were obtained is mentioned. For example, DenseNet-201 (round 4, minibatch size 10) means that when the DenseNet-201 network was trained for the 4th time (independently) using the same training and test data, precision/sensitivity/recall of 0.68/0.63/0.85 was obtained for SEK.\u003c/p\u003e\n\u003cp\u003eTo understand features that contributed to the classification, Grad-CAM was applied to all test images (Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e). The 1100\u0026ndash;1200 1/cm and 1350\u0026ndash;1450 1/cm regions were responsible for 10 out of 19 and 4 out of 19 ACK spectra image classification (Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003ea). In 5 images, sections in the background seems to have been responsible for ACK classification. 10 out of 19 BCC image classification were due to 1600\u0026ndash;1700 1/cm region, 14 of 19 were due to 1100\u0026ndash;1200 1/cm region, and none were due to background sections (Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003eb). 12 out of 17 SEK images were classified due to 1600\u0026ndash;1700 1/cm region while 4 of 17 images were classified due to 1100\u0026ndash;1200 1/cm.\u003c/p\u003e\n\u003cp\u003eWhen these spectral regions were plotted as subplots, saved as images (Fig. \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e) and used as input for DenseNet-201 (mini batch size 8, round 5), improvement in sensitivity (0.78) for BCC was observed (Table \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e, strategy 3) compared to using entire range images as input (specificity \u0026ndash; 0.7) (Table \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e, strategy 4). When the same spectral ranges plotted as subplots \u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003ewithout axes\u003c/span\u003e (Fig. \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e), saved as images, and used as input for GoogLeNet (round 10, mini batch size 10), improvement in precision (0.76) and specificity (0.89) was observed (Table \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e, method d)) compared to method b (precision/specificity \u0026ndash; 0.68/0.83). Note that precision and specificity of ACK also increased to 1, but sensitivity was low (0.41).\u003c/p\u003e\n\u003cp\u003e\u003c/p\u003e\u0026nbsp;\u003ctable id=\"Tab1\" border=\"1\" style=\"margin-right: calc(0%); width: 100%;\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eClassification using different strategies 1) using spectra intensity values as input for traditional machine learning (ML) methods, 2. using spectra graphs saved as images as input for pretrained image networks, 3. using spectral regions identified by Grad-CAM mapping plotted as subplots with axes and saved as images as input for pretrained image networks, and 4. using spectral regions identified by Grad-CAM mapping plotted as subplots without axes and saved as images as input for pretrained image networks. SVM: Support Vector Machine. NB: Na\u0026iuml;ve Bayes. Opt SVM: SVM model with hyperparameters optimized.\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" rowspan=\"2\" style=\"width: 4.477%;\"\u003e\n \u003cp\u003eMethod\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" rowspan=\"2\" style=\"width: 7.1952%;\"\u003e\n \u003cp\u003eML/DL\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"3\" style=\"width: 8.7941%;\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"3\" style=\"width: 8.7941%;\"\u003e\n \u003cp\u003eSensitivity\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"4\" style=\"width: 10.3131%;\"\u003e\n \u003cp\u003eSpecificity/Recall\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"4\" style=\"width: 8.7941%;\"\u003e\n \u003cp\u003eF-measure\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"4\" style=\"width: 14.3904%;\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 2.1809%;\"\u003e\n \u003cp\u003eRange (1/cm)\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003eACK\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003eBCC\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003eSEK\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003eACK\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003eBCC\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003eSEK\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003eACK\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003eBCC\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003eSEK\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003eACK\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003eBCC\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003eSEK\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003eACK\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003eBCC\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003eSEK\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\" style=\"width: 14.3105%;\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" rowspan=\"5\" style=\"width: 4.477%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.1952%;\"\u003e\n \u003cp\u003eSVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.82\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.54\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.59\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.52\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.63\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.94\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.64\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.61\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.61\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.62\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.62\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.62\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" rowspan=\"8\" style=\"width: 14.3105%;\"\u003e\n \u003cp\u003e900\u0026ndash;1800\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 7.1952%;\"\u003e\n \u003cp\u003eNB\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.56\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.53\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.56\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.67\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.74\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.65\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.61\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.63\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.28\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.54\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.54\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.54\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 7.1952%;\"\u003e\n \u003cp\u003eOpt SVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.57\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.63\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.61\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.59\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.70\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.52\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.80\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.83\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.58\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.67\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.56\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.60\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.60\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.60\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 7.1952%;\"\u003e\n \u003cp\u003eSVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.82\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.54\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.59\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.52\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.63\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.78\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.64\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.61\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.61\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.62\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.62\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.62\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 7.1952%;\"\u003e\n \u003cp\u003eOpt SVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.57\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.63\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.61\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.59\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.52\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.83\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.58\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.67\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.56\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.60\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.60\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.60\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" rowspan=\"3\" style=\"width: 4.477%;\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.1952%;\"\u003e\n \u003cp\u003eDenseNet-201\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.79\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.63\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.52\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.70\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.56\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.63\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.91\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.83\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.75\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.59\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.57\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.63\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.63\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.63\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 7.1952%;\"\u003e\n \u003cp\u003eGoogLeNet\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.68\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.59\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.48\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.70\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.59\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.76\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.83\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.49\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.69\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.59\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.59\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.59\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.59\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 7.1952%;\"\u003e\n \u003cp\u003eDenseNet-201\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.68\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.61\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.68\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.63\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.63\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.85\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.85\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.65\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.66\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.65\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.65\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e0.65\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.65\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 4.477%;\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.1952%;\"\u003e\n \u003cp\u003eDenseNet-201\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.59\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.66\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.63\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.78\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.37\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.8\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.81\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.61\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.71\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.43\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.59\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.59\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.59\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 14.3105%;\"\u003e\n \u003cp\u003e1100\u0026ndash;1200, 1350\u0026ndash;1450,1600\u0026ndash;1700\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 4.477%;\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.1952%;\"\u003e\n \u003cp\u003eGoogLeNet\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.76\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.56\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.41\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.70\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.89\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.63\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.58\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.73\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.69\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.038%;\"\u003e\n \u003cp\u003e0.68\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.958%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.68\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 2.7981%;\"\u003e\n \u003cp\u003e0.68\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 14.3105%;\"\u003e\n \u003cp\u003e1100\u0026ndash;1200, 1350\u0026ndash;1450,1600\u0026ndash;1700\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003c/p\u003e\n"},{"header":"Discussion","content":"\u003cp\u003eA search of keywords \u0026ldquo;spectroscopy medical diagnosis classification\u0026rdquo; in Google Scholar revealed roughly 180 articles since 2025 in the first 20 pages itself, pertaining to use of biospectroscopy for medical disease diagnosis using machine learning (ML)-based classification. From experience of working with biospectroscopy groups, many more articles are not submitted or fail to get published because they do not achieve good classification results. Especially with multi-class classification problems, where number of groups exceed six or ten, getting good classification for all classes get tougher. A new method of classifying spectra is hence required.\u003c/p\u003e \u003cp\u003ePretrained image networks are trained to classify images belong to thousand or more different groups and are well-suited to tackle multi-class classification problems with large number of groups. Woo et.al. have used images of graphs of multi-sensor signals for classification of sleep stages\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e. Williams et. al. used of GoogLeNet and ResNet for classification of microplastics Raman and FTIR spectra using images of spectrum graphs\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eIn the current study, we explored three novel aspects of using spectral graph images as input for pretrained image networks \u0026ndash; 1. whether this strategy is better than traditional ML methods such as SVM, LDA, and so on, 2. whether this strategy can give us the features used by networks for classification, and 3. whether these features can be selectively used to improve classification. We found that pretrained image networks give better classification than traditional ML, and that spectral regions responsible for classification can be identified and used to improve classification. The study was done using a small dataset, hence there can be a question on whether the results will be true for large datasets. However, most biomedical researchers start with small datasets to explore study feasibility, and poor classification tend to discourage researchers from moving to larger studies. Hence, testing on small dataset may be more apt.\u003c/p\u003e \u003cp\u003eIt should be noted that one method did not give good classification for all groups. SVM without any optimization gave the best classification for ACK, while SVM post optimization gave best results for BCC and SEK. Similarly, DenseNet-201 gave the best results for ACK and SEK, while GoogLeNet gave the best results for BCC. We envision using different trained models in parallel for classification in real-world applications. For example, to identify whether an unknown sample is ACK, BCC, or SEK, both DenseNet-201 and GoogLeNet will be used for prediction. If the predictions do not match, weightage will be given to the network best trained for predicted group. For example, if DenseNet predicts ACK, and GoogLeNet predicts SEK, the verdict will be ACK as DenseNet predicts ACK better.\u003c/p\u003e \u003cp\u003eRegarding feature identification, pretrained image networks have an edge over traditional ML. PCA loading factors or Shapley values used to identify features for ML methods give underlying features in general, but not for each spectrum. With pretrained networks, Grad-CAM mapping allows feature identification in each spectrum image (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). In Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea, some ACK images are classified due to 1100\u0026ndash;1200 1/cm region while others were due to 1350\u0026ndash;1450 1/cm. This allows better understanding of the basis of classification. Further, it is observed that some images have been classified based on background. Their predictions can be marked as \u0026ldquo;suspect\u0026rdquo;. During decision making, as in the example in the previous paragraph, if DenseNet predicts ACK, but basis of classification is image background, while GoogLeNet predicts SEK, where the basis of classification is the spectral region, then verdict will be SEK.\u003c/p\u003e \u003cp\u003eOnce spectral regions are identified, graphs of regions plotted as subplots could be used to improve sensitivity/specificity. There are multiple ways to do this, one of which was explored in the study \u0026ndash; subplots without axes. Other options are to arrange subplots in horizontal manner instead of vertically. Each region may be further magnified by sub plotting sections. For example, instead of subplots of 1100\u0026ndash;1200 and 1350\u0026ndash;1450 1/cm, subplots of 1100\u0026ndash;1150, 1150\u0026ndash;1200, 1350\u0026ndash;1400, and 1400\u0026ndash;1450 1/cm can be generated and used as input. MATLAB allows \u0026ldquo;no spacing option\u0026rdquo; in \u0026ldquo;tiledlayout\u0026rdquo; to remove space between subplots. Simple algorithms can be generated to further remove white space, if any.\u003c/p\u003e \u003cp\u003eAnother advantage of pretrained networks is that they can be trained multiple times. The results of training vary. This allows the best trained model to be chosen; the best model being the one that gives the highest precision/sensitivity/specificity and where most or all images have been classified based on spectral region instead of background. In this study, each network was trained 20 times to choose the best trained model. Further, training options can be varied. Stochastic Gradient Descent with Momentum (sgdm), Root Mean Square Propagation, or Adaptive Moment Estimation, etc. can be used as solvers. In this study, only \u0026lsquo;sgdm\u0026rsquo; was used, but other methods can be explored to improve classification. The mini batch size can be varied. In the current study, bath sizes 8, 10, and 12 were used, each giving different results. While the initial learn rate was kept 0.00001 in this study, this can be varied to improve results.\u003c/p\u003e \u003cp\u003eFinally, the process can be fully automated. We developed a MATLAB script to automatically import and preprocess spectra, plot spectrum/spectrum ranges/subplots as per user input, save these as images in automatically generated labelled subfolders, import images with label information into image datastores, train and test different pretrained networks, save trained models, and collate classification outcomes in an excel sheet. The code can be made available on request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eMartin FL (2023) Translating Biospectroscopy Techniques to Clinical Settings: A New Paradigm in Point-of-Care Screening and/or Diagnostics. J Personalised Med 13:1511\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKannan S, Callery EL, Rowbottom AW (2023) Vibrational biospectroscopy in the clinical setting: Exploring the impact of new advances in the field of immunology. Journal of Spectroscopy. 5557441 (2023)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim Y, Bui TT, Chung H (2025) A review on diverse spectroscopic methods for identification of hepato-pancreato-biliary diseases using human bile as a specimen. Appl Spectrosc Rev, 1\u0026ndash;20\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSon Y et al (2025) Advances in spectroscopic detection of traumatic brain injury biomarkers: Potential for diagnostic applications. Appl Spectrosc Rev, 1\u0026ndash;30\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDelrue C, De Bruyne S, Speeckaert MM (2025) The Promise of Infrared Spectroscopy in Liquid Biopsies for Solid Cancer Detection. Diagnostics 15:368\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXuesong H et al (2024) Commentary on the review articles of spectroscopy technology combined with chemometrics in the last three years. Appl Spectrosc Rev 59:423\u0026ndash;482\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSpola\u0026ocirc;r N et al (2024) Fine-tuning pre-trained neural networks for medical image classification in small clinical datasets. Multimedia Tools Appl 83:27305\u0026ndash;27329\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eArchana R, Jeevaraj PE (2024) Deep learning models for digital image processing: a review. Artif Intell Rev 57:11\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRaghavan K (2024) Attention guided grad-CAM: an improved explainable artificial intelligence model for infrared breast cancer detection. Multimedia Tools Appl 83:57551\u0026ndash;57578\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eda Cunha PHP et al (2027) NIR-SC-UFES: A portable NIR spectral dataset to skin cancer in vivo. medRxiv, 2024.2011. 24317165 (2024)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBertolini E, Precision, Specificity, Sensitivity A (2025) \u0026amp; F1-score (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.mathworks.com/matlabcentral/fileexchange/86158-precision-specificity-sensitivity-accuracy-f1-score)\u003c/span\u003e\u003cspan address=\"https://www.mathworks.com/matlabcentral/fileexchange/86158-precision-specificity-sensitivity-accuracy-f1-score)\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. MATLAB Central File Exchange. Retrieved April 24, (2025)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWoo Y et al (2023) Automatic Sleep Stage Classification Using Deep Learning Algorithm for Multi-Institutional Database. IEEE Access 11:46297\u0026ndash;46307\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWilliams WA, Arunprasad A, Aravamudhan S (2024) Application of a modified set of GoogLeNet and ResNet-18 convolutional neural networks towards the identification of environmentally derived-MPLs in the Yadkin-pee dee river basin. Environ Syst Res 13:53\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-6562812/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6562812/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe study compares sensitivity/specificity of classification by pretrained image networks and traditional Machine Learning (ML) methods. One hundred seven spectra each of benign skin conditions actinic keratosis (ACK) and seborrheic keratosis (SEK), and skin cancer basal cell carcinoma (BCC) were downloaded from a public database. Eighty spectra per group were used for training and twenty-seven spectra per group for testing. In the first classification strategy, spectrum intensity values were used as input for Linear Discriminant Analysis (LDA), K-Nearest Neighbors (KNN), Decision Tree (DT), TreeBagger, Ensemble method, Na\u0026iuml;ve Bayes, Support Vector Machine (SVM), and Artificial Neural Network (ANN). The second strategy involved using spectrum graphs saved as images to train GoogLeNet, Places-365 GoogLeNet, ResNet-50, Inception-V3, DenseNet-201, and NasNetMobile. Strategy 2 yielded better sensitivity/specificity \u0026ndash; 0.7/ 0.91 (ACK), 0.7/0.83 (BCC), and 0.63/0.85 (SEK) compared to strategy 1\u0026ndash;0.52/0.94 (ACK), 0.7/0.8 (BCC), and 0.5/0.8 (SEK). Grad-CAM mapping suggested that 1100\u0026ndash;1200,1350\u0026ndash;1450, and 1600\u0026ndash;1700 1/cm to be responsible for classification by strategy 2. When these regions plotted as subplots and saved as images were used for training using strategy 2, sensitivity for BCC increases to 0.78. Results suggest using pretrained image networks to classify spectra may yield better results, give a visual understanding of the basis of classification, and provide means to improve classification further.\u003c/p\u003e","manuscriptTitle":"GoogLeNet/DenseNet-201 to classify near-infrared (NIR) spectrum graphs for cancer diagnosis – using pretrained image networks for medical spectroscopy","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-06 13:11:25","doi":"10.21203/rs.3.rs-6562812/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"328b3f32-69a2-43cc-b819-473a6b962d54","owner":[],"postedDate":"May 6th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":48011089,"name":"Physical sciences/Optics and photonics/Other photonics/Biophotonics"},{"id":48011090,"name":"Physical sciences/Optics and photonics/Optical techniques/Optical spectroscopy/Near-infrared spectroscopy"}],"tags":[],"updatedAt":"2025-05-06T13:11:25+00:00","versionOfRecord":[],"versionCreatedAt":"2025-05-06 13:11:25","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6562812","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6562812","identity":"rs-6562812","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}