{"paper_id":"337c9f38-10e1-4dd5-81fe-144669dfbfff","body_text":"Few-Shot Learning using Siamese Twin Network for the Classification of Blood Cells | 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 Research Article Few-Shot Learning using Siamese Twin Network for the Classification of Blood Cells Sudhakar Tummala, Anil K Suresh This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-2215631/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 Automated classification of blood cells from microscopic images is an interesting research area owing to advancements of efficient neural network architectures. Here, we developed a few-shot contrastive learning model for the classification of peripheral blood cells including lymphocytes, monocytes, basophils, eosinophils, neutrophils, immature granulocytes, erythroblasts, and platelets using EfficientNet as a base model and contrastive loss as a loss function. A total of 17092 publicly accessible images acquired using the CellaVision DM96 were analyzed. From 125 images of each cell type, 20000 image pairs are generated for Siamese twin network (STN) training and another 125 images from each cell type are used for few-shot validation. Therefore, out of 17092 images, 6% were used for training, 6% for few-shot validation and rest 88% for few-shot testing. This architecture demonstrates an overall accuracy of 97.21% during 8-way 3-shot testing for the classification of all cell types with an accuracy of 97.72% for the classification of white blood cells alone. Further, we propose a novel class activation mapping scheme for the interpretability of the model decisions suitable for STN. To conclude, the proposed framework based on contrastive learning could be used for the fully automated self-exploratory classification and identification of peripheral blood cells. Siamese network EfficientNet few-shot learning blood cell typing explainability Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction Automated classification of peripheral blood cells ( e.g ., basophils, neutrophils, eosinophils, monocytes, lymphocytes, immature granulocytes, erythroblasts, and platelets) is a prime requisite for medical researchers. As, it provides hassle-free support to cytologists and pathologists in analyzing blood smears under normal and disease states [ 1 , 2 ]. Additionally, understanding the structural integrity of peripheral blood cells is crucial for the identification and monitoring of several diseases including hematologic conditions [ 3 ]. With the success of AlexNet on the ImageNet classification challenge [ 4 ], advancements in computational infrastructure, and parameter optimization methods, researchers across the globe are developing novel neural network architectures for various computer vision tasks. Prominent successful models after AlexNet are VGGNets, ResNets, DenseNets, MobileNets, Inception, Xception networks, and EfficientNets that not only outperform on a specific computer vision task, but the knowledge gained by them can be used to tackle other computer vision tasks with the advent of transfer learning approaches [ 5 – 7 ]. Previous studies for the classification of normal peripheral blood cell images employed few of the aforementioned state-of-the-art architectures and their variants as base models using transfer learning [ 3 , 8 , 9 ]. Studies also investigated the feasibility of deep learning models for the exclusive classification of white blood cells [ 10 – 13 ]. However, all these methods used multitudes of images for training leaving behind only few images for testing. Typically, large neural networks demand huge data to achieve better generalizability, and in medical imaging, data availability is a scarce and it is time consuming to collect large-scale data with quality acquaintance. Lately, few-shot learning (FSL) based methods are gaining tremendous attention as they relatively require less data for model training and yet are robust and efficient. The FSL based methods can be trained via contrastive learning (CL) of the input data either in pairs or triplets or quadruplets and by performing quantitative comparisons of their embeddings and representations [ 14 – 16 ]. If the embeddings are adjacent in the latent space, they belong to the same class and vice versa. Therefore, the objective of CL is to make the embeddings of the same class nearer and embeddings of dissimilar classes farther in the latent space under some constraints [ 17 , 18 ]. Hence, Siamese twin network (STN) can be constructed for CL as few recent studies have successfully employed these networks for several medical imaging investigations [ 19 – 23 ]. Several methods are proposed to highlight various important regions in the image for model decision in a classification task such as class activation mapping (CAM) [ 24 ], gradCAM [ 25 ], and gradCAM++ [ 26 ]. CAM uses the trained weights between the final convolution layer and the dense layer as the weights to be multiplied with the final activation maps. Whereas gradCAM uses the gradients obtained during backpropagation as weights which is the generalized version of CAM and gradCAM + + and general form of gradCAM. All these methods are extensively applied in single feed forward network architectures for visual saliency mappings. Whereas, in STN models, the saliency mapping methods are sparse and adapt already existing gradCAM methods [ 27 , 28 ]. EfficientNets have been a versatile neural network architectures that outperformed the aforementioned novel architectures, and the network has relatively few parameters. It also exhibits better speed during various image recognition tasks and performs satisfactorily on transfer learning diverse datasets [ 29 ]. EfficientNets work on more principled compound scaling of width, depth of the network, and resolution of the input images while maintaining model efficiency. Moreover, the visual saliency maps generated from EfficientNets show that the model focuses on more relevant object regions [ 29 ]. To the best of our knowledge, FSL with deep contrastive representation learning is not yet been explored for multi-labeled classification of normal peripheral blood cells along with visual saliency mappings for interpretability of the model predictions. Therefore, herein, we conducted the following: (a) Developed an STN with EfficientNet as the base model and contrastive loss as a loss function for the classification of normal peripheral blood cells, (b) Eight-way few-shot verification for the identification of specific cell image type from the eight types of peripheral blood cells. The few-shot verification involves few-shot validation for hyperparameter tuning and few-shot testing, and eventually, c) Developed a novel visual saliency mapping scheme without necessitating gradients computation for explanation of the proposed STN decisions during few-shot verification. Methods This section describes the dataset of peripheral blood cell images, Siamese twin network training with EfficientNet as the base model, N-way few-shot validation and testing, performance metrics, and generation of class activation maps between query image and support set. Table 1 Summary of the cell types in the dataset, number of images for each cell type, number of images used for Siamese twin network training, few-shot validation, and few-shot testing. N: number of images. Blood Cell Type Training (N) Validation (N) Testing (N) Total (N) Basophils 125 125 968 1218 Eosinophils 125 125 2867 3117 Neutrophils 125 125 3079 3329 Lymphocytes 125 125 964 1214 Monocytes 125 125 1170 1420 Immature Granulocytes 125 125 2645 2895 Erythroblasts 125 125 1301 1551 Platelets 125 125 2098 2348 Total 1000 1000 15092 17092 Dataset In the current work, an openly accessible dataset of normal peripheral blood cells was utilized from the Hospital Clinic of Barcelona. The dataset contains 17092 RGB images that were captured using the CellaVision DM96 [ 30 ]. The images of different types of blood cells include neutrophils, basophils, eosinophils, lymphocytes, monocytes, immature granulocytes, erythroblast, and platelets. Predominant image resolution was 360 × 363 with very few images having the resolution of 360 × 360 or 359 × 360. The images were graded with pre-determined cell types by expert clinical pathologists from the same clinic and were used as ground truth labels. Complete details about the dataset are given in Table 1 . Sample images (one image per class) from the dataset are shown in Fig. 1 . Generation of Image Pairs We have considered 125 images from each class with a total of 1000 images for training. Twenty image pairs are created for each image cumulating to a total of 20000 image pairs for STN training. When both the paired images belong to the same cell type it is termed as a positive pair labelled as 0 and if the images are dissimilar, they are designated as a negative pair labelled as 1. To avoid imbalance in the classes, 10000 positive pairs and 10000 negative pairs are randomly produced. Out of the 20 generated pairs for each image, ten were same pairs and ten were different pairs. To maintain uniformity in differently paired, we ensured that each specific cell was paired with rest of the seven cell types at least once. Before feeding the image pairs to the STN model, the images were initially verified to ensure that their intensity range is between zero and 255, a primary requirement for EfficientNets. Siamese Neural Network Architecture Figure 2 shows the proposed STN architecture with EfficientNet-B3 as the base model. The final softmax layer of the base model was discarded and a global average pooling layer is incorporated. The EfficientNet-B3 model was employed here to transform the input sample from image space to embedding space with a mapping \\(\\varphi (.)\\) . For the input image pair \\({X}_{1}\\) and \\({X}_{2}\\) , \\({\\varphi (X}_{1})\\) and \\({\\varphi (X}_{2})\\) are their embeddings in the latent space respectively. Since the goal of the STN model is to make the embeddings of similar pairs closer and vice versa, the quantitative comparison of the embeddings via absolute differences is implemented using the lambda layer. After which, a sigmoid neuron is placed leading to a probability between zero and one, where a value less than 0.5 indicates a positive pair and vice versa. EfficientNet and Tuning of Hyperparameters Since the softmax layer of the EfficientNet-B3 model was eliminated, it outputs a feature tensor of size 10×10×1536 which is also the output for the final convolution layer. Further, this output feature tensor is global averaged to achieve a feature vector of the size 1536. Furthermore, the 1536 feature tensor is connected to the lambda layer for its quantitative comparison with another feature tensor of length 1536 as described earlier. The base model approximately contains twenty million parameters and therefore to reduce the computational time as well as to leverage the power of transfer learning, selected parameters (weights and biases) of the final ten percent layers from the EfficientNet-B3 model were allowed to get updated during backpropagation and rest of the model parameters were nontrainable. There are several hyperparameters in the proposed model that can be tuned such as mini-batch size, learning rate, number of epochs, and the choice of optimizer during gradient descent. We evaluated models with different possible combinations with respect to four adaptive gradient descent optimizers namely RMSprop [ 31 ], Adam [ 32 ], Nadam [ 33 ], Adadelta [ 34 ]. Finally, for few-shot testing, we selected the model that gave better accuracies during few-shot validation as described in Table 2 . Contrastive Training For contrastive training, NVIDIA Tesla P100 GPU with 26 GB RAM available in Google Colab Pro is utilized. It has a TensorFlow backend with Keras API. As required by EfficientNet-B3, the images are center cropped to attain a resolution of 300×300. Further, the loss metric for updating the model parameters through backpropagation is contrastive loss ( \\({C}_{l})\\) that is computed using Eq. (1) given below: $${ C}_{l}=\\left(1-y\\right)*{\\sigma \\left(d\\left(\\varphi \\left({X}_{1}\\right), {\\varphi (X}_{2})\\right)\\right)}^{2}+y*{\\left\\{\\text{max}\\left(0, m-\\sigma \\left(d\\left(\\varphi \\left({X}_{1}\\right), \\varphi \\left({X}_{2}\\right)\\right)\\right)\\right)\\right\\}}^{2} \\left(1\\right)$$ where \\(d\\left(\\varphi \\right({X}_{1}), \\varphi ({X}_{2}\\left)\\right)\\) is the distance metric and in this study, it is the weighted sum of the absolute differences between the embeddings \\({\\varphi (X}_{1})\\) and \\({\\varphi (X}_{2})\\) . Above, \\(y\\) indicates the true label of the image pair, m is the distance margin that is set to one in the current study and \\(\\sigma\\) is the sigmoid function as described in Eq. (2). $$\\sigma \\left(d\\left(.\\right)\\right)=\\frac{1}{1+{e}^{-d(.)}} \\left(2\\right)$$ Let \\({a}_{d}\\) is a tensor represents the absolute differences between the embedding tensors as given in Eq. (3). $${a}_{d}^{i}=\\left|{\\varphi (X}_{1}^{i})-{\\varphi (X}_{2}^{i})\\right| \\forall i \\left(3\\right)$$ Eventually, the distance metric \\(d\\left(\\varphi \\right({X}_{1}), \\varphi ({X}_{2}\\left)\\right)\\) is represented using the expression in Eq. (4). $$d\\left(\\varphi \\left({X}_{1}\\right), \\varphi \\left({X}_{2}\\right)\\right)=\\sum _{i=1}^{N}{w}_{i}{a}_{d}^{i}+{b}_{i} \\left(4\\right)$$ In Equations (3) and (4), \\(\\varphi \\left({X}_{1}^{i}\\right)\\) and \\({\\varphi (X}_{2}^{i})\\) are the i th values of the embedding vectors \\({\\varphi (X}_{1})\\) and \\({\\varphi (X}_{2})\\) respectively, \\(N\\) is the length of the tensor \\({a}_{d}\\) and \\({w}_{i}\\) and \\({b}_{i}\\) are the weights and biases that needs to be learned during backpropagation. Support Set A support set that is necessary for few-shot validation and testing containing eight images is formed from the images in the test set. To represent each class in the support set, one image is randomly sampled from the images of the corresponding class for N -way k -shot validation and testing as detailed below. Whenever k is greater than one, we used entirely new support set for each shot. N -way k -Shot Validation and Testing The value of N in N -way is the number of classes which is set to eight in the present study and k in k -shot is between one to five for few-shot learning. If k is equal to one, it is called one-shot learning and if k is equal to two, it is two-shot learning, and so on. In this study, we performed 8-way 1-shot, 8-way 2-shot, 8-way 3-shot, and 8-way 5-shot validation and testing for multiclass classification of eight cell types. Finally, the class of the query image is decided based on the highest similarity with respect to the images in the support set. Mathematically, the class prediction for a query image \\({x}_{q}\\) for 8-way k -shot learning is given in Eq. (5). $${Y}_{q} = argmin\\left\\{\\sum _{i=1}^{k}{\\Psi }{({X}_{q}, {X}_{s})}_{i}\\right\\} \\left(5\\right)$$ Above, \\({X}_{q}=\\left\\{{x}_{q}^{c}={x}_{q}:1\\le c\\le N \\right\\}\\) , that means \\({X}_{q}\\) is the set of images where the query image \\({x}_{q}\\) is repeated N times to match with the number of images in the support set \\({X}_{s}=\\left\\{{x}_{s}^{c}:1\\le c\\le N\\right\\}\\) so that the comparison of \\({x}_{q}\\) with \\({X}_{s}\\) would happen in one epoch, \\({\\Psi }\\left({X}_{q}, {X}_{s}\\right)\\) is the prediction result of STN model which is a vector of N similarity values between \\({X}_{q}\\) and \\({X}_{s}\\) . Finally, \\({Y}_{q}\\) is the predicted label for the query image \\({x}_{q}\\) that is between zero and seven. The calculation of overall accuracy for predicting a class c during k -shot validation/testing is implemented using expression (6), where \\({N}_{c}^{{\\prime }}\\) is correctly predicted images for calss c and \\({N}_{c}\\) is the total number of images in class c . $$Overall Accuracy = \\frac{\\sum _{i=1}^{k}({\\frac{{N}_{c}^{{\\prime }}}{{N}_{c}})}_{i}}{k} \\left(6\\right)$$ Creation of Visual Saliency Maps For the creation of saliency maps, to infer explainability to the network while making decisions, the output of the lambda layer of the STN is used as the weight tensor. The activation maps \\({A}^{H\\times W\\times C}\\) of the final convolution layer of the EfficientNet-B3 for the query image are multiplied with the weight vector to get weighted activation maps \\({A}_{w}^{H\\times W\\times C}\\) as described by Eq. (7). $${A}_{w}^{H\\times W\\times i}= {A}^{H\\times W\\times i}* {a}_{d}^{i} \\forall i \\left(7\\right)$$ In Eq. (5), \\({a}_{d}^{i}\\) is the weight tensor as already described in Eq. (3). Afterward, an average activation map \\({A}_{m}^{H\\times W}\\) of spatial size of \\(H\\times W\\) is obtained by averaging all \\(C\\) weighted activation maps as described in Eq. (8). $${A}_{m}^{H\\times W}=ReLU\\left(\\frac{1}{C}\\sum _{i=1}^{C}{A}_{w}^{H\\times W\\times i} \\right) \\left(8\\right)$$ The negative values in the mean activation map are removed using the \\(ReLU\\) (rectified linear unit) activation function, which is given in Eq. (9). $$ReLU\\left(z\\right)= \\left\\{\\begin{array}{c}z if z>0\\\\ 0 if z\\le 0\\end{array}\\right. \\left(9\\right)$$ For EfficientNet-B3, \\(H\\times W\\times C\\) = \\(10\\times 10\\times 1536\\) . The depth of the activation maps \\(C\\) and length of the weight tensor \\(N\\) are identical. Finally, the 10×10 coarse activation map \\({A}_{m}^{H\\times W}\\) is resized to match with the spatial resolution of the input query image which is 300×300 using python based scikit-image toolbox. For identification of highly activated regions in the query image when it is compared with the images in the support set, the resized heatmaps are overlaid onto the corresponding RGB images in the support set showing the most similar/dissimilar regions. Results The 8-way few-shot validation with the overall accuracies of the proposed STN architecture using EfficientNet-B3 as the base model are given in Table 2 . For different combinations of mini batch gradient descent optimizers and other model hyperparameters where Adadelta performed relatively better are ascribed. Hence, the selected hyperparameters for few-shot testing along with Adadelta optimizer are mini-batch size: 16, learning rate: 1.0, number of epochs: 15. The overall validation accuracy values are in the range of 94.7%-97.1%. The 8-way few-shot test overall accuracies are given in Fig. 6 and the best overall test accuracy over 97.2% is obtained when k is equal to 3. The few-shot test overall accuracy for the classification of white blood cells alone is 97.7%. The classification value is highest for platelets with 99.8%. Table 3 shows the results for overall test accuracies obtained in comparison with the previous studies. The number of images used for training, validation, and testing are also mentioned for a fair comparison. The saliency maps highlighting the most discriminative/similar regions between the query image and the images in the support sets for basophil and erythroblast cells are given in Fig. 3 and Fig. 4 respectively. The 1D tensor \\({a}_{d}\\) is reshaped into a 2D tensor of size 48 × 32 and is also plotted along with the activation maps. Further, Fig. 5 shows the saliency maps highlighting the regions for each cell type comparatively with another image of the same cell type in the support set to understand the reason for their similarity. As anticipated, the regions where most of the representative features located for decision making are highlighted and those regions are cell bodies and their surroundings. Table 2 Different optimizers and hyperparameters along with accuracy values during Siamese network training and overall accuracy values during 8-way 1-shot, 2-shot, 3-shot, and 5-shot validation on 1000 images with 125 samples from each class. SGD : stochastic gradient descent, lr : learning rate, ne : number of epochs, mbs : mini batch size. Optimizer & Hyperparameters Accuracy Siamese training 8-way 1-shot validation 8-way 2-shot validation 8-way 3-shot validation 8-way 5-shot validation \\(RMSprop\\) \\(\\left\\{\\begin{array}{c}lr=0.001, ne=15, mbs=16\\\\ lr=0.001, ne=15, mbs=32\\\\ lr=0.0005, ne=20, mbs=64\\end{array}\\right.\\) \\(\\left\\{\\begin{array}{c}99.80\\\\ 99.89\\\\ 99.93\\end{array}\\right.\\) \\(\\left\\{\\begin{array}{c}95.31\\\\ 96.10\\\\ 94.63\\end{array}\\right.\\) \\(\\left\\{\\begin{array}{c}96.40\\\\ 96.31\\\\ 95.50\\end{array}\\right.\\) \\(\\left\\{\\begin{array}{c}96.40\\\\ 96.50\\\\ 95.60\\end{array}\\right.\\) \\(\\left\\{\\begin{array}{c}96.51\\\\ 96.60\\\\ 95.60\\end{array}\\right.\\) \\(Adam\\) \\(\\left\\{\\begin{array}{c}lr=0.001, ne=15, mbs=16\\\\ lr=0.001, ne=15, mbs=32\\\\ lr=0.0005, ne=20, mbs=64\\end{array}\\right.\\) \\(\\left\\{\\begin{array}{c}99.74\\\\ 99.79\\\\ 99.91\\end{array}\\right.\\) \\(\\left\\{\\begin{array}{c}95.30\\\\ 93.10\\\\ 95.20\\end{array}\\right.\\) \\(\\left\\{\\begin{array}{c}96.80\\\\ 95.35\\\\ 96.30\\end{array}\\right.\\) \\(\\left\\{\\begin{array}{c}96.82\\\\ 95.40\\\\ 96.50\\end{array}\\right.\\) \\(\\left\\{\\begin{array}{c}97.01\\\\ 95.40\\\\ 96.60\\end{array}\\right.\\) \\(Nadam\\) \\(\\left\\{\\begin{array}{c}lr=0.001, ne=15, mbs=16\\\\ lr=0.01, ne=15, mbs=32\\\\ lr=0.01, ne=20, mbs=64\\end{array}\\right.\\) \\(\\left\\{\\begin{array}{c}99.84\\\\ 99.81\\\\ 99.74\\end{array}\\right.\\) \\(\\left\\{\\begin{array}{c}94.70\\\\ 93.00\\\\ 93.70\\end{array}\\right.\\) \\(\\left\\{\\begin{array}{c}96.20\\\\ 95.70\\\\ 94.60\\end{array}\\right.\\) \\(\\left\\{\\begin{array}{c}95.70\\\\ 95.40\\\\ 96.00\\end{array}\\right.\\) \\(\\left\\{\\begin{array}{c}96.25\\\\ 95.32\\\\ 95.40\\end{array}\\right.\\) \\(Adadelta\\) \\(\\left\\{\\begin{array}{c}lr=1.0, ne=15, mbs=16\\\\ lr=1.0, ne=15, mbs=32\\\\ lr=0.9, ne=20, mbs=64\\end{array}\\right.\\) \\(\\left\\{\\begin{array}{c}99.97\\\\ 99.94\\\\ 99.95\\end{array}\\right.\\) \\(\\left\\{\\begin{array}{c}96.10\\\\ 95.80\\\\ 96.00\\end{array}\\right.\\) \\(\\left\\{\\begin{array}{c}97.01\\\\ 96.90\\\\ 96.90\\end{array}\\right.\\) \\(\\left\\{\\begin{array}{c}96.90\\\\ 96.50\\\\ 97.10\\end{array}\\right.\\) \\(\\left\\{\\begin{array}{c}97.03\\\\ 97.00\\\\ 97.00\\end{array}\\right.\\) Discussion In this study, we assessed multi label classification of normal peripheral blood cell images that include white blood cells, immature granulocytes, erythroblasts, and platelets using STN architecture based on contrastive learning. Extensive evaluations for the choice of the optimizer and other model hyperparameters were performed that suggested the optimal choice based on their performance during few-shot validation. In general, the performance scores of the models given in Table 2 and were determined to be decent irrespective of the optimizer and other hyperparameters, with Adadelta optimizer as a slightly better variation. As expected, based on \\({a}_{d}\\) plots from Figs. 3 and 4 , it can be inferred that the embeddings are closer for similar pairs and apart for dissimilar pairs. Our proposed method performed better [ 3 ] or slightly worse [ 9 ] compared to previous studies that were tested on the same dataset. Even though the overall few-shot testing accuracy is two percent less than the test accuracies from the reported study [ 8 ], our methodology is completely different where we have utilized only 6% of the data for training to further propel the feasibility of CL based models in the medical domain. As also discussed in [ 3 ], due to the nucleus shape and structure similarities among igs, basophils, neutrophils, and monocytes, there is a higher percent of misclassifications among these cell types. From Fig. 6 , we can notice that 2.27%, 2.53% of igs are wrongly classified as monocytes and neutrophils respectively. Similarly, 2.48% of basophils are misclassified as igs. The classification accuracy for platelets is 99.8% since their size and morphology are entirely different compared to other cell types. In terms of overall test accuracy for the classification of only white blood cells, our results are comparable with several studies that used deep transfer learning approaches or much complicated neural network designs [ 10 – 13 ]. Our study achieved an accuracy of 97.7% for white blood cell classification which is better than some of the reported accuracies 96.1% and 96.9% respectively [ 10 , 11 ]. Our overall accuracy is comparable with the reported average accuracy of 98.8% in a recent study that used Siamese network [ 35 ], however, that study was restricted to white blood cells, and it used only 430 images for testing and moreover, no visual saliency mappings is implemented for explainability. Table 3 Comparison of our overall test accuracy in percentage, number of images used for training, validation and testing with previously reported studies. N: number of images. Study Test Accuracy (%) Training (N) Validation (N) Testing (N) Acevedo et.al., 2019 [ 3 ] 96.20 11077 4096 1919 Ucar, 2020 [ 9 ] 97.94 13674 1709 1709 Long et.al., 2021 [ 8 ] 99.30 13674 - 3418 Our study 97.21 1000 1000 15092 Going beyond the accuracy values for comparison of current studies with previous works, and even though we have used 20000 generated image pairs, they are obtained just using 125 images from each class and yet achieved comparable test performance metrics both while classifying all cell types and white blood cells exclusively. After extensive few-shot validation, hyperparameters set based on the Adadelta optimizer performed well and in addition, it can essentially eliminate learning rate selection by fixing it to one. Further, as expected, the 8-way k -shot validation and testing values are better when k is greater than 2 since the greater number of images per class in the support set could lead to greater classification performance. To our knowledge, none of the previous studies listed in Table 3 based on traditional CNN based classification had explored the class activation mapping using either novel methods or already existing schemes such as CAM [ 24 ], gradCAM [ 25 ], and gradCAM++ [ 36 ]. Our proposed saliency mapping method is simple as it does not require gradients computation unlike gradCAM. Since we adapted the \\({a}_{d}\\) tensor values as the weights to be multiplied with the final convolution layer activation maps of the query image, explicit retraining of the model with the query image and the support set is not essential. Moreover, from Figures (3), (4), and (5), we could use the proposed saliency mapping method for cell localization in the image, and eventually, the framework may be adapted for coarse level cell segmentation. In the future, the proposed framework could be tested on blood cell images of disease conditions such as anemia and leukemia. Conclusions We proposed a contrastive learning based approach that relatively require fewer samples for training with finetuned EfficientNet as the base model for classification of peripheral blood cell images with the overall few-shot test accuracy of 97.21%. We employed state-of-the-art EfficientNet-B3 deep learning model and finetuned the last ten percent layers to learn more representative features of blood cells. In our 8-way few-shot experiments, the Adadelta optimizer with the choice of other hyperparameters provided slightly better results with the highest accuracy being 99.8% for classifying platelets. Three-shot and 5-shot testing accuracies are superior to 1-shot and 2-shot testing. Further, we demonstrated the creation of visual saliency maps without computing the gradients to interpret the decision of the Siamese network. Hence, our procedure may facilitate as a reference for deep CL based peripheral blood cell image classification against which the future research can be compared. In conclusion, we believe that our methodology could facilitate to development of deep learning based methods for classification of peripheral blood cells with fewer data samples and with interpretability. Declarations Competing Interests: The authors have no competing interests to declare. Funding: No funding was received for conducting this study. Compliance with Ethical Standards This research study was conducted retrospectively using human subject data made available in open access by Hospital Clinic Barcelona. Ethical approval was not required as confirmed by the license attached with the open access data. Acknowledgements We would like to thank Hospital Clinic Barcelona for providing the dataset and SRM University—AP for providing the research infrastructure. References Ceelie H, Dinkelaar RB, van Gelder W (2007) Examination of peripheral blood films using automated microscopy; evaluation of Diffmaster Octavia and Cellavision DM96. Journal of Clinical Pathology 60:72. https://doi.org/10.1136/JCP.2005.035402 Rümke CL (1985) Imprecision of ratio-derived differential leukocyte counts. Blood Cells 11:311–4, 315 Acevedo A, Alférez S, Merino A, et al (2019) Recognition of peripheral blood cell images using convolutional neural networks. Computer Methods and Programs in Biomedicine 180:105020. https://doi.org/10.1016/J.CMPB.2019.105020 Krizhevsky A, Sutskever I, Hinton GE (2012) ImageNet Classification with Deep Convolutional Neural Networks. Advances in Neural Information Processing Systems 25: Tan C, Sun F, Kong T, et al (2018) A Survey on Deep Transfer Learning. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 11141 LNCS:270–279. https://doi.org/10.1007/978-3-030-01424-7_27 Pathak Y, Shukla PK, Tiwari A, et al (2020) Deep Transfer Learning Based Classification Model for COVID-19 Disease. IRBM. https://doi.org/10.1016/J.IRBM.2020.05.003 Minaee S, Kafieh R, Sonka M, et al (2020) Deep-COVID: Predicting COVID-19 from chest X-ray images using deep transfer learning. Medical Image Analysis 65:101794. https://doi.org/10.1016/J.MEDIA.2020.101794 Long F, Peng JJ, Song W, et al (2021) BloodCaps: A capsule network based model for the multiclassification of human peripheral blood cells. Computer methods and programs in biomedicine 202:. https://doi.org/10.1016/J.CMPB.2021.105972 Ucar F (2020) Deep learning approach to cell classification in human peripheral blood. 5th International Conference on Computer Science and Engineering, UBMK 2020 383–387. https://doi.org/10.1109/UBMK50275.2020.9219480 Baydilli YY, Atila Ü (2020) Classification of white blood cells using capsule networks. Computerized medical imaging and graphics: the official journal of the Computerized Medical Imaging Society 80:. https://doi.org/10.1016/J.COMPMEDIMAG.2020.101699 Shahin AI, Guo Y, Amin KM, Sharawi AA (2019) White blood cells identification system based on convolutional deep neural learning networks. Computer methods and programs in biomedicine 168:69–80. https://doi.org/10.1016/J.CMPB.2017.11.015 Kutlu H, Avci E, Özyurt F (2020) White blood cells detection and classification based on regional convolutional neural networks. Medical hypotheses 135:. https://doi.org/10.1016/J.MEHY.2019.109472 Almezhghwi K, Serte S (2020) Improved Classification of White Blood Cells with the Generative Adversarial Network and Deep Convolutional Neural Network. Computational Intelligence and Neuroscience 2020:. https://doi.org/10.1155/2020/6490479 Medela A, Picon A, Saratxaga CL, et al (2019) Few shot learning in histopathological images: Reducing the need of labeled data on biological datasets. Proceedings - International Symposium on Biomedical Imaging 2019-April:1860–1864. https://doi.org/10.1109/ISBI.2019.8759182 Puch S, Sánchez I, Rowe M (2019) Few-Shot Learning with Deep Triplet Networks for Brain Imaging Modality Recognition. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 11795 LNCS:181–189. https://doi.org/10.1007/978-3-030-33391-1_21 Chen X, Yao L, Zhou T, et al (2021) Momentum contrastive learning for few-shot COVID-19 diagnosis from chest CT images. Pattern Recognition 113:107826. https://doi.org/10.1016/J.PATCOG.2021.107826 Chopra S, Hadsell R, LeCun Y (2005) Learning a similarity metric discriminatively, with application to face verification. Proceedings – 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, CVPR 2005 I:539–546. https://doi.org/10.1109/CVPR.2005.202 Weinberger KQ, Saul LK (2009) Distance Metric Learning for Large Margin Nearest Neighbor Classification. Journal of Machine Learning Research 10:207–244. https://doi.org/10.5555/1577069 Tummala S (2021) Deep Learning Framework using Siamese Neural Network for Diagnosis of Autism from Brain Magnetic Resonance Imaging. In: 2021 6th International Conference for Convergence in Technology (I2CT). IEEE, pp 1–5 Madhu G, Lalith Bharadwaj B, Rohit B, et al (2021) Convolutional Siamese networks for one-shot malaria parasite recognition in microscopic images. Demystifying Big Data, Machine Learning, and Deep Learning for Healthcare Analytics 277–306. https://doi.org/10.1016/B978-0-12-821633-0.00007-6 Rossi A, Hosseinzadeh M, Bianchini M, et al (2021) Multi-Modal Siamese Network for Diagnostically Similar Lesion Retrieval in Prostate MRI. IEEE Transactions on Medical Imaging 40:986–995. https://doi.org/10.1109/TMI.2020.3043641 Li MD, Chang K, Bearce B, et al (2020) Siamese neural networks for continuous disease severity evaluation and change detection in medical imaging. npj Digital Medicine 2020 3:1 3:1–9. https://doi.org/10.1038/s41746-020-0255-1 Wang J, Fang Z, Lang N, et al (2017) A multi-resolution approach for spinal metastasis detection using deep Siamese neural networks. Computers in Biology and Medicine 84:137–146. https://doi.org/10.1016/J.COMPBIOMED.2017.03.024 Zhou B, Khosla A, Lapedriza A, et al (2015) Learning Deep Features for Discriminative Localization. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition 2016-December:2921–2929. https://doi.org/10.1109/CVPR.2016.319 Selvaraju RR, Cogswell M, Das A, et al (2016) Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization. International Journal of Computer Vision 128:336–359. https://doi.org/10.1007/s11263-019-01228-7 Chattopadhyay A, Sarkar A, Howlader P, Balasubramanian VN (2017) Grad-CAM++: Improved Visual Explanations for Deep Convolutional Networks. Proceedings – 2018 IEEE Winter Conference on Applications of Computer Vision, WACV 2018 2018-January:839–847. https://doi.org/10.1109/WACV.2018.00097 Yoo TK, Choi JY, Kim HK (2021) Feasibility study to improve deep learning in OCT diagnosis of rare retinal diseases with few-shot classification. Medical and Biological Engineering and Computing 59:401–415. https://doi.org/10.1007/S11517-021-02321-1/FIGURES/12 Chen L, Chen J, Hajimirsadeghi H, Mori G (2020) Adapting Grad-CAM for Embedding Networks. Proceedings – 2020 IEEE Winter Conference on Applications of Computer Vision, WACV 2020 2783–2792. https://doi.org/10.1109/WACV45572.2020.9093461 Tan M, Le Q V. (2019) EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. 36th International Conference on Machine Learning, ICML 2019 2019-June:10691–10700 Acevedo A, Merino A, Alférez S, et al (2020) A dataset of microscopic peripheral blood cell images for development of automatic recognition systems. Data in Brief 30:105474. https://doi.org/10.1016/J.DIB.2020.105474 Dauphin YN, De Vries H, Bengio Y (2015) RMSProp and equilibrated adaptive learning rates for non-convex optimization. Advances in Neural Information Processing Systems 2015-January:1504–1512 Kingma DP, Ba JL (2014) Adam: A Method for Stochastic Optimization. 3rd International Conference on Learning Representations, ICLR 2015 - Conference Track Proceedings Dozat T (2015) Incorporating Nesterov Momentum into Adam Zeiler MD (2012) ADADELTA: An Adaptive Learning Rate Method Wang Y, Cao Y (2020) A computer-assisted human peripheral blood leukocyte image classification method based on Siamese network. Medical & Biological Engineering & Computing 2020 58:7 58:1575–1582. https://doi.org/10.1007/S11517-020-02180-2 Chattopadhyay A, Sarkar A, Howlader P, Balasubramanian VN (2017) Grad-CAM++: Improved Visual Explanations for Deep Convolutional Networks. Proceedings – 2018 IEEE Winter Conference on Applications of Computer Vision, WACV 2018 2018-January:839–847. https://doi.org/10.1109/WACV.2018.00097 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-2215631\",\"acceptedTermsAndConditions\":true,\"allowDirectSubmit\":true,\"archivedVersions\":[],\"articleType\":\"Research Article\",\"associatedPublications\":[],\"authors\":[{\"id\":147847537,\"identity\":\"bfe0b105-a2ac-4d29-a0c9-eba16460d1d0\",\"order_by\":0,\"name\":\"Sudhakar Tummala\",\"email\":\"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA8ElEQVRIiWNgGAWjYHACMwaGAyCa+QBMxACMCGthY0sgWQsPXBl+9fIzkrc9+HHGLp9Bvufz58ocuzwG9uZtEgwFd3BqMbiRVm7YcyPZsoGNd5vk2W3JxQw8x8okGAye4dYikWMmwfOB2YABqIWxcRtzYgNIhMHgMB6H5ZhJ/vlQD9TC8/hj47b6xAb5N/i1MNzIMZPmuXEYpIVBsnHbYaAtPPi1GJx5ViYtc+a4ARtbmhlQy/FiNp60YosEfA5rT94m+eZYtQE/82GQw6rz+NkPb7zx4Q8eh8EAG5ROADMSCGtAAJIUj4JRMApGwcgAAJMUTi8OJDZ0AAAAAElFTkSuQmCC\",\"orcid\":\"https://orcid.org/0000-0001-5735-9418\",\"institution\":\"SRM University AP\",\"correspondingAuthor\":true,\"prefix\":\"\",\"firstName\":\"Sudhakar\",\"middleName\":\"\",\"lastName\":\"Tummala\",\"suffix\":\"\"},{\"id\":147847538,\"identity\":\"179f5701-cb95-43c3-a16b-4c31a34c1c3f\",\"order_by\":1,\"name\":\"Anil K Suresh\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"SRM University AP\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Anil\",\"middleName\":\"K\",\"lastName\":\"Suresh\",\"suffix\":\"\"}],\"badges\":[],\"createdAt\":\"2022-10-29 08:16:31\",\"currentVersionCode\":1,\"declarations\":{\"humanSubjects\":true,\"vertebrateSubjects\":false,\"conflictsOfInterestStatement\":true,\"humanSubjectEthicalGuidelines\":true,\"humanSubjectConsent\":true,\"humanSubjectClinicalTrial\":false,\"humanSubjectCaseReport\":false,\"vertebrateSubjectEthicalGuidelines\":false,\"coiExplicitlySet\":false},\"doi\":\"10.21203/rs.3.rs-2215631/v1\",\"doiUrl\":\"https://doi.org/10.21203/rs.3.rs-2215631/v1\",\"draftVersion\":[],\"editorialEvents\":[],\"editorialNote\":\"\",\"failedWorkflow\":false,\"files\":[{\"id\":28475184,\"identity\":\"2ca54729-eed1-4f11-a771-04e0200962ac\",\"added_by\":\"auto\",\"created_at\":\"2022-10-31 20:01:04\",\"extension\":\"jpg\",\"order_by\":1,\"title\":\"Figure 1\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":59222,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003eMicroscopy images of different blood cells as per their class randomly extracted from the clinical dataset. a) basophil, b) eosinophil, c) neutrophil, d) lymphocyte, e) monocyte, f) immature granulocyte, g) erythroblast, and h) platelet.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"Fig1.jpg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-2215631/v1/b0f779e2f4332f0c75bfe50a.jpg\"},{\"id\":28475187,\"identity\":\"185a64d9-5d5b-4cd0-bb6c-4ebd426f8d3a\",\"added_by\":\"auto\",\"created_at\":\"2022-10-31 20:01:04\",\"extension\":\"jpg\",\"order_by\":2,\"title\":\"Figure 2\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":60645,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003eProposed Siamese twin network architecture with EfficientNet-B3 as the base model for contrastive training. \\u003cem\\u003eϕ\\u003c/em\\u003e(\\u003cem\\u003eX\\u003c/em\\u003e1)\\u0026nbsp;is the embedding of image \\u003cem\\u003eX\\u003c/em\\u003e1\\u0026nbsp;and \\u003cem\\u003eϕ\\u003c/em\\u003e(\\u003cem\\u003eX\\u003c/em\\u003e2)\\u0026nbsp;is the embedding of image \\u003cem\\u003eX\\u003c/em\\u003e2 GAP: global average pooling, σ\\u0026nbsp;: sigmoid neuron.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"Fig2.jpg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-2215631/v1/5cdcfb1eff4fc0f13e33a2c7.jpg\"},{\"id\":28475189,\"identity\":\"700154fc-c244-4b6d-b680-ff8a7d398517\",\"added_by\":\"auto\",\"created_at\":\"2022-10-31 20:01:04\",\"extension\":\"jpg\",\"order_by\":3,\"title\":\"Figure 3\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":86135,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003eReshaped values of absolute differences between the embeddings (\\u003cem\\u003ead\\u003c/em\\u003e) and most activated regions during one-shot testing of a basophil query image. These regions are overlaid onto the images in the support set. The differences are lower when the basophil query image is compared with the basophil support image and the differences are higher when the query image is compared with non-basophil images in the support set. a: basophil, b: eosinophil, c: immature granulocyte, d: erythroblast, e: lymphocyte, f: monocyte, g: platelet, h: neutrophil.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"Fig3.jpg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-2215631/v1/9501356654825c35f58265e8.jpg\"},{\"id\":28475532,\"identity\":\"3df5cc0d-8a77-4745-bbae-7ea994b1276e\",\"added_by\":\"auto\",\"created_at\":\"2022-10-31 20:06:04\",\"extension\":\"jpg\",\"order_by\":4,\"title\":\"Figure 4\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":95566,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003eReshaped values of absolute differences between the embeddings (\\u003cem\\u003ead\\u003c/em\\u003e) and most discriminative regions during one-shot testing of an erythroblast query image. The discriminative regions are overlaid onto the images in the support set. The differences are lower when the erythroblast query image is compared with the erythroblast support image and differences are higher when the query image is compared with non-basophil images in the support set. a: basophil, b: eosinophil, c: immature granulocyte, d: erythroblast, e: lymphocyte, f: monocyte, g: platelet, h: neutrophil.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"Fig4.jpg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-2215631/v1/8177a5a13583b9393b7cb14b.jpg\"},{\"id\":28475186,\"identity\":\"c8891893-2d18-4ea1-87a5-287a45660098\",\"added_by\":\"auto\",\"created_at\":\"2022-10-31 20:01:04\",\"extension\":\"jpg\",\"order_by\":5,\"title\":\"Figure 5\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":81858,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003eQuery cell image and the corresponding same cell image in the support set with highlighted regions that contributed to the most similarity between the two images during one-shot testing. a: basophil, b: eosinophil, c: immature granulocyte, d: erythroblast, e: lymphocyte, f: monocyte, g: platelet, h: neutrophil.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"Fig5.jpg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-2215631/v1/f3958010d95d15f57642b56f.jpg\"},{\"id\":28475190,\"identity\":\"f4b5268e-7c4e-46ee-b51b-00fdb19ab4e8\",\"added_by\":\"auto\",\"created_at\":\"2022-10-31 20:01:04\",\"extension\":\"jpg\",\"order_by\":6,\"title\":\"Figure 6\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":96585,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003eConfusion matrices with accuracy values along the diagonal during 8-way few-shot testing on 15092 samples. Off diagonal values indicate percent of misclassifications. Top row is for 1-shot and 2-shot testing and bottom row is for 3-shot, and 5-shot testing.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"Fig6.jpg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-2215631/v1/709a9f3792d2484a01416c31.jpg\"},{\"id\":28475538,\"identity\":\"ffc66dd2-8752-4109-b40f-7bedc1e0b919\",\"added_by\":\"auto\",\"created_at\":\"2022-10-31 20:06:10\",\"extension\":\"pdf\",\"order_by\":0,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"manuscript-pdf\",\"size\":799857,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"manuscript.pdf\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-2215631/v1/692966be-6d45-4d67-afa4-7e19954938d1.pdf\"}],\"financialInterests\":\"\",\"formattedTitle\":\"\\u003cp\\u003e\\u003cstrong\\u003eFew-Shot Learning using Siamese Twin Network for the Classification of Blood Cells\\u003c/strong\\u003e\\u003c/p\\u003e\",\"fulltext\":[{\"header\":\"Introduction\",\"content\":\"\\u003cp\\u003eAutomated classification of peripheral blood cells (\\u003cem\\u003ee.g\\u003c/em\\u003e., basophils, neutrophils, eosinophils, monocytes, lymphocytes, immature granulocytes, erythroblasts, and platelets) is a prime requisite for medical researchers. As, it provides hassle-free support to cytologists and pathologists in analyzing blood smears under normal and disease states [\\u003cspan citationid=\\\"CR1\\\" class=\\\"CitationRef\\\"\\u003e1\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR2\\\" class=\\\"CitationRef\\\"\\u003e2\\u003c/span\\u003e]. Additionally, understanding the structural integrity of peripheral blood cells is crucial for the identification and monitoring of several diseases including hematologic conditions [\\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e]. With the success of AlexNet on the ImageNet classification challenge [\\u003cspan citationid=\\\"CR4\\\" class=\\\"CitationRef\\\"\\u003e4\\u003c/span\\u003e], advancements in computational infrastructure, and parameter optimization methods, researchers across the globe are developing novel neural network architectures for various computer vision tasks. Prominent successful models after AlexNet are VGGNets, ResNets, DenseNets, MobileNets, Inception, Xception networks, and EfficientNets that not only outperform on a specific computer vision task, but the knowledge gained by them can be used to tackle other computer vision tasks with the advent of transfer learning approaches [\\u003cspan additionalcitationids=\\\"CR6\\\" citationid=\\\"CR5\\\" class=\\\"CitationRef\\\"\\u003e5\\u003c/span\\u003e\\u0026ndash;\\u003cspan citationid=\\\"CR7\\\" class=\\\"CitationRef\\\"\\u003e7\\u003c/span\\u003e].\\u003c/p\\u003e \\u003cp\\u003ePrevious studies for the classification of normal peripheral blood cell images employed few of the aforementioned state-of-the-art architectures and their variants as base models using transfer learning [\\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR8\\\" class=\\\"CitationRef\\\"\\u003e8\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR9\\\" class=\\\"CitationRef\\\"\\u003e9\\u003c/span\\u003e]. Studies also investigated the feasibility of deep learning models for the exclusive classification of white blood cells [\\u003cspan additionalcitationids=\\\"CR11 CR12\\\" citationid=\\\"CR10\\\" class=\\\"CitationRef\\\"\\u003e10\\u003c/span\\u003e\\u0026ndash;\\u003cspan citationid=\\\"CR13\\\" class=\\\"CitationRef\\\"\\u003e13\\u003c/span\\u003e]. However, all these methods used multitudes of images for training leaving behind only few images for testing. Typically, large neural networks demand huge data to achieve better generalizability, and in medical imaging, data availability is a scarce and it is time consuming to collect large-scale data with quality acquaintance. Lately, few-shot learning (FSL) based methods are gaining tremendous attention as they relatively require less data for model training and yet are robust and efficient. The FSL based methods can be trained \\u003cem\\u003evia\\u003c/em\\u003e contrastive learning (CL) of the input data either in pairs or triplets or quadruplets and by performing quantitative comparisons of their embeddings and representations [\\u003cspan additionalcitationids=\\\"CR15\\\" citationid=\\\"CR14\\\" class=\\\"CitationRef\\\"\\u003e14\\u003c/span\\u003e\\u0026ndash;\\u003cspan citationid=\\\"CR16\\\" class=\\\"CitationRef\\\"\\u003e16\\u003c/span\\u003e]. If the embeddings are adjacent in the latent space, they belong to the same class and vice versa. Therefore, the objective of CL is to make the embeddings of the same class nearer and embeddings of dissimilar classes farther in the latent space under some constraints [\\u003cspan citationid=\\\"CR17\\\" class=\\\"CitationRef\\\"\\u003e17\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR18\\\" class=\\\"CitationRef\\\"\\u003e18\\u003c/span\\u003e]. Hence, Siamese twin network (STN) can be constructed for CL as few recent studies have successfully employed these networks for several medical imaging investigations [\\u003cspan additionalcitationids=\\\"CR20 CR21 CR22\\\" citationid=\\\"CR19\\\" class=\\\"CitationRef\\\"\\u003e19\\u003c/span\\u003e\\u0026ndash;\\u003cspan citationid=\\\"CR23\\\" class=\\\"CitationRef\\\"\\u003e23\\u003c/span\\u003e].\\u003c/p\\u003e \\u003cp\\u003eSeveral methods are proposed to highlight various important regions in the image for model decision in a classification task such as class activation mapping (CAM) [\\u003cspan citationid=\\\"CR24\\\" class=\\\"CitationRef\\\"\\u003e24\\u003c/span\\u003e], gradCAM [\\u003cspan citationid=\\\"CR25\\\" class=\\\"CitationRef\\\"\\u003e25\\u003c/span\\u003e], and gradCAM++ [\\u003cspan citationid=\\\"CR26\\\" class=\\\"CitationRef\\\"\\u003e26\\u003c/span\\u003e]. CAM uses the trained weights between the final convolution layer and the dense layer as the weights to be multiplied with the final activation maps. Whereas gradCAM uses the gradients obtained during backpropagation as weights which is the generalized version of CAM and gradCAM\\u0026thinsp;+\\u0026thinsp;+\\u0026thinsp;and general form of gradCAM. All these methods are extensively applied in single feed forward network architectures for visual saliency mappings. Whereas, in STN models, the saliency mapping methods are sparse and adapt already existing gradCAM methods [\\u003cspan citationid=\\\"CR27\\\" class=\\\"CitationRef\\\"\\u003e27\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR28\\\" class=\\\"CitationRef\\\"\\u003e28\\u003c/span\\u003e]. EfficientNets have been a versatile neural network architectures that outperformed the aforementioned novel architectures, and the network has relatively few parameters. It also exhibits better speed during various image recognition tasks and performs satisfactorily on transfer learning diverse datasets [\\u003cspan citationid=\\\"CR29\\\" class=\\\"CitationRef\\\"\\u003e29\\u003c/span\\u003e]. EfficientNets work on more principled compound scaling of width, depth of the network, and resolution of the input images while maintaining model efficiency. Moreover, the visual saliency maps generated from EfficientNets show that the model focuses on more relevant object regions [\\u003cspan citationid=\\\"CR29\\\" class=\\\"CitationRef\\\"\\u003e29\\u003c/span\\u003e]. To the best of our knowledge, FSL with deep contrastive representation learning is not yet been explored for multi-labeled classification of normal peripheral blood cells along with visual saliency mappings for interpretability of the model predictions.\\u003c/p\\u003e \\u003cp\\u003eTherefore, herein, we conducted the following: (a) Developed an STN with EfficientNet as the base model and contrastive loss as a loss function for the classification of normal peripheral blood cells, (b) Eight-way few-shot verification for the identification of specific cell image type from the eight types of peripheral blood cells. The few-shot verification involves few-shot validation for hyperparameter tuning and few-shot testing, and eventually, c) Developed a novel visual saliency mapping scheme without necessitating gradients computation for explanation of the proposed STN decisions during few-shot verification.\\u003c/p\\u003e\"},{\"header\":\"Methods\",\"content\":\"\\u003cp\\u003eThis section describes the dataset of peripheral blood cell images, Siamese twin network training with EfficientNet as the base model, N-way few-shot validation and testing, performance metrics, and generation of class activation maps between query image and support set.\\u003c/p\\u003e \\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\\u003eSummary of the cell types in the dataset, number of images for each cell type, number of images used for Siamese twin network training, few-shot validation, and few-shot testing. N: number of images.\\u003c/p\\u003e \\u003c/div\\u003e \\u003c/caption\\u003e \\u003ccolgroup cols=\\\"5\\\"\\u003e \\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c1\\\" colnum=\\\"1\\\"\\u003e\\u003c/div\\u003e \\u003cdiv align=\\\"char\\\" char=\\\".\\\" class=\\\"colspec\\\" colname=\\\"c2\\\" colnum=\\\"2\\\"\\u003e\\u003c/div\\u003e \\u003cdiv align=\\\"char\\\" char=\\\".\\\" 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 \\u003cthead\\u003e \\u003ctr\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003eBlood Cell Type\\u003c/p\\u003e \\u003c/th\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003eTraining (N)\\u003c/p\\u003e \\u003c/th\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003eValidation (N)\\u003c/p\\u003e \\u003c/th\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003eTesting (N)\\u003c/p\\u003e \\u003c/th\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003eTotal (N)\\u003c/p\\u003e \\u003c/th\\u003e \\u003c/tr\\u003e \\u003c/thead\\u003e \\u003ctbody\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003eBasophils\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e125\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e125\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e968\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e1218\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003eEosinophils\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e125\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e125\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e2867\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e3117\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003eNeutrophils\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e125\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e125\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e3079\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e3329\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003eLymphocytes\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e125\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e125\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e964\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e1214\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003eMonocytes\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e125\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e125\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e1170\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e1420\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003eImmature Granulocytes\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e125\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e125\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e2645\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e2895\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003eErythroblasts\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e125\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e125\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e1301\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e1551\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003ePlatelets\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e125\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e125\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e2098\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e2348\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003e\\u003cb\\u003eTotal\\u003c/b\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e\\u003cb\\u003e1000\\u003c/b\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e\\u003cb\\u003e1000\\u003c/b\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e\\u003cb\\u003e15092\\u003c/b\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e\\u003cb\\u003e17092\\u003c/b\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003c/tbody\\u003e \\u003c/colgroup\\u003e \\u003c/table\\u003e\\u003c/div\\u003e \\u003c/p\\u003e \\u003cdiv id=\\\"Sec3\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003eDataset\\u003c/h2\\u003e \\u003cp\\u003eIn the current work, an openly accessible dataset of normal peripheral blood cells was utilized from the Hospital Clinic of Barcelona. The dataset contains 17092 RGB images that were captured using the CellaVision DM96 [\\u003cspan citationid=\\\"CR30\\\" class=\\\"CitationRef\\\"\\u003e30\\u003c/span\\u003e]. The images of different types of blood cells include neutrophils, basophils, eosinophils, lymphocytes, monocytes, immature granulocytes, erythroblast, and platelets. Predominant image resolution was 360 \\u0026times; 363 with very few images having the resolution of 360 \\u0026times; 360 or 359 \\u0026times; 360. The images were graded with pre-determined cell types by expert clinical pathologists from the same clinic and were used as ground truth labels. Complete details about the dataset are given in Table\\u0026nbsp;\\u003cspan refid=\\\"Tab1\\\" class=\\\"InternalRef\\\"\\u003e1\\u003c/span\\u003e. Sample images (one image per class) from the dataset are shown in Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig1\\\" class=\\\"InternalRef\\\"\\u003e1\\u003c/span\\u003e.\\u003c/p\\u003e \\u003c/div\\u003e \\u003cdiv id=\\\"Sec4\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003eGeneration of Image Pairs\\u003c/h2\\u003e \\u003cp\\u003eWe have considered 125 images from each class with a total of 1000 images for training. Twenty image pairs are created for each image cumulating to a total of 20000 image pairs for STN training. When both the paired images belong to the same cell type it is termed as a positive pair labelled as 0 and if the images are dissimilar, they are designated as a negative pair labelled as 1. To avoid imbalance in the classes, 10000 positive pairs and 10000 negative pairs are randomly produced. Out of the 20 generated pairs for each image, ten were same pairs and ten were different pairs. To maintain uniformity in differently paired, we ensured that each specific cell was paired with rest of the seven cell types at least once. Before feeding the image pairs to the STN model, the images were initially verified to ensure that their intensity range is between zero and 255, a primary requirement for EfficientNets.\\u003c/p\\u003e \\u003c/div\\u003e \\u003cdiv id=\\\"Sec5\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003eSiamese Neural Network Architecture\\u003c/h2\\u003e \\u003cp\\u003eFigure\\u0026nbsp;\\u003cspan refid=\\\"Fig2\\\" class=\\\"InternalRef\\\"\\u003e2\\u003c/span\\u003e shows the proposed STN architecture with EfficientNet-B3 as the base model. The final \\u003cem\\u003esoftmax\\u003c/em\\u003e layer of the base model was discarded and a global average pooling layer is incorporated. The EfficientNet-B3 model was employed here to transform the input sample from image space to embedding space with a mapping \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\varphi (.)\\\\)\\u003c/span\\u003e\\u003c/span\\u003e. For the input image pair \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({X}_{1}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e and \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({X}_{2}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e, \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({\\\\varphi (X}_{1})\\\\)\\u003c/span\\u003e\\u003c/span\\u003e and \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({\\\\varphi (X}_{2})\\\\)\\u003c/span\\u003e\\u003c/span\\u003e are their embeddings in the latent space respectively. Since the goal of the STN model is to make the embeddings of similar pairs closer and vice versa, the quantitative comparison of the embeddings \\u003cem\\u003evia\\u003c/em\\u003e absolute differences is implemented using the \\u003cem\\u003elambda\\u003c/em\\u003e layer. After which, a sigmoid neuron is placed leading to a probability between zero and one, where a value less than 0.5 indicates a positive pair and vice versa.\\u003c/p\\u003e\\u003c/div\\u003e \\u003cdiv id=\\\"Sec6\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003eEfficientNet and Tuning of Hyperparameters\\u003c/h2\\u003e \\u003cp\\u003eSince the softmax layer of the EfficientNet-B3 model was eliminated, it outputs a feature tensor of size 10\\u0026times;10\\u0026times;1536 which is also the output for the final convolution layer. Further, this output feature tensor is global averaged to achieve a feature vector of the size 1536. Furthermore, the 1536 feature tensor is connected to the \\u003cem\\u003elambda\\u003c/em\\u003e layer for its quantitative comparison with another feature tensor of length 1536 as described earlier. The base model approximately contains twenty million parameters and therefore to reduce the computational time as well as to leverage the power of transfer learning, selected parameters (weights and biases) of the final ten percent layers from the EfficientNet-B3 model were allowed to get updated during backpropagation and rest of the model parameters were nontrainable.\\u003c/p\\u003e \\u003cp\\u003eThere are several hyperparameters in the proposed model that can be tuned such as mini-batch size, learning rate, number of epochs, and the choice of optimizer during gradient descent. We evaluated models with different possible combinations with respect to four adaptive gradient descent optimizers namely \\u003cem\\u003eRMSprop\\u003c/em\\u003e [\\u003cspan citationid=\\\"CR31\\\" class=\\\"CitationRef\\\"\\u003e31\\u003c/span\\u003e], \\u003cem\\u003eAdam\\u003c/em\\u003e [\\u003cspan citationid=\\\"CR32\\\" class=\\\"CitationRef\\\"\\u003e32\\u003c/span\\u003e], \\u003cem\\u003eNadam\\u003c/em\\u003e [\\u003cspan citationid=\\\"CR33\\\" class=\\\"CitationRef\\\"\\u003e33\\u003c/span\\u003e], \\u003cem\\u003eAdadelta\\u003c/em\\u003e [\\u003cspan citationid=\\\"CR34\\\" class=\\\"CitationRef\\\"\\u003e34\\u003c/span\\u003e]. Finally, for few-shot testing, we selected the model that gave better accuracies during few-shot validation as described in Table\\u0026nbsp;\\u003cspan refid=\\\"Tab2\\\" class=\\\"InternalRef\\\"\\u003e2\\u003c/span\\u003e.\\u003c/p\\u003e \\u003c/div\\u003e \\u003cdiv id=\\\"Sec7\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003eContrastive Training\\u003c/h2\\u003e \\u003cp\\u003eFor contrastive training, NVIDIA Tesla P100 GPU with 26 GB RAM available in Google Colab Pro is utilized. It has a TensorFlow backend with Keras API. As required by EfficientNet-B3, the images are center cropped to attain a resolution of 300\\u0026times;300. Further, the loss metric for updating the model parameters through backpropagation is contrastive loss (\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({C}_{l})\\\\)\\u003c/span\\u003e\\u003c/span\\u003e that is computed using Eq.\\u0026nbsp;(1) given below:\\u003cdiv id=\\\"Equa\\\" class=\\\"Equation\\\"\\u003e\\u003cdiv format=\\\"TEX\\\" class=\\\"mathdisplay\\\" id=\\\"FileID_Equa\\\" name=\\\"EquationSource\\\"\\u003e\\n$${ C}_{l}=\\\\left(1-y\\\\right)*{\\\\sigma \\\\left(d\\\\left(\\\\varphi \\\\left({X}_{1}\\\\right), {\\\\varphi (X}_{2})\\\\right)\\\\right)}^{2}+y*{\\\\left\\\\{\\\\text{max}\\\\left(0, m-\\\\sigma \\\\left(d\\\\left(\\\\varphi \\\\left({X}_{1}\\\\right), \\\\varphi \\\\left({X}_{2}\\\\right)\\\\right)\\\\right)\\\\right)\\\\right\\\\}}^{2} \\\\left(1\\\\right)$$\\u003c/div\\u003e\\u003c/div\\u003e\\u003c/p\\u003e \\u003cp\\u003ewhere \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(d\\\\left(\\\\varphi \\\\right({X}_{1}), \\\\varphi ({X}_{2}\\\\left)\\\\right)\\\\)\\u003c/span\\u003e\\u003c/span\\u003e is the distance metric and in this study, it is the weighted sum of the absolute differences between the embeddings \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({\\\\varphi (X}_{1})\\\\)\\u003c/span\\u003e\\u003c/span\\u003e and \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({\\\\varphi (X}_{2})\\\\)\\u003c/span\\u003e\\u003c/span\\u003e. Above, \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(y\\\\)\\u003c/span\\u003e\\u003c/span\\u003e indicates the true label of the image pair, \\u003cem\\u003em\\u003c/em\\u003e is the distance margin that is set to one in the current study and \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\sigma\\\\)\\u003c/span\\u003e\\u003c/span\\u003e is the sigmoid function as described in Eq.\\u0026nbsp;(2).\\u003cdiv id=\\\"Equb\\\" class=\\\"Equation\\\"\\u003e\\u003cdiv format=\\\"TEX\\\" class=\\\"mathdisplay\\\" id=\\\"FileID_Equb\\\" name=\\\"EquationSource\\\"\\u003e\\n$$\\\\sigma \\\\left(d\\\\left(.\\\\right)\\\\right)=\\\\frac{1}{1+{e}^{-d(.)}} \\\\left(2\\\\right)$$\\u003c/div\\u003e\\u003c/div\\u003e\\u003c/p\\u003e \\u003cp\\u003eLet \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({a}_{d}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e is a tensor represents the absolute differences between the embedding tensors as given in Eq.\\u0026nbsp;(3).\\u003cdiv id=\\\"Equc\\\" class=\\\"Equation\\\"\\u003e\\u003cdiv format=\\\"TEX\\\" class=\\\"mathdisplay\\\" id=\\\"FileID_Equc\\\" name=\\\"EquationSource\\\"\\u003e\\n$${a}_{d}^{i}=\\\\left|{\\\\varphi (X}_{1}^{i})-{\\\\varphi (X}_{2}^{i})\\\\right| \\\\forall i \\\\left(3\\\\right)$$\\u003c/div\\u003e\\u003c/div\\u003e\\u003c/p\\u003e \\u003cp\\u003eEventually, the distance metric \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(d\\\\left(\\\\varphi \\\\right({X}_{1}), \\\\varphi ({X}_{2}\\\\left)\\\\right)\\\\)\\u003c/span\\u003e\\u003c/span\\u003e is represented using the expression in Eq.\\u0026nbsp;(4).\\u003cdiv id=\\\"Equd\\\" class=\\\"Equation\\\"\\u003e\\u003cdiv format=\\\"TEX\\\" class=\\\"mathdisplay\\\" id=\\\"FileID_Equd\\\" name=\\\"EquationSource\\\"\\u003e\\n$$d\\\\left(\\\\varphi \\\\left({X}_{1}\\\\right), \\\\varphi \\\\left({X}_{2}\\\\right)\\\\right)=\\\\sum _{i=1}^{N}{w}_{i}{a}_{d}^{i}+{b}_{i} \\\\left(4\\\\right)$$\\u003c/div\\u003e\\u003c/div\\u003e\\u003c/p\\u003e \\u003cp\\u003eIn Equations (3) and (4), \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\varphi \\\\left({X}_{1}^{i}\\\\right)\\\\)\\u003c/span\\u003e\\u003c/span\\u003e and \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({\\\\varphi (X}_{2}^{i})\\\\)\\u003c/span\\u003e\\u003c/span\\u003e are the \\u003cem\\u003ei\\u003c/em\\u003e\\u003csup\\u003eth\\u003c/sup\\u003e values of the embedding vectors \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({\\\\varphi (X}_{1})\\\\)\\u003c/span\\u003e\\u003c/span\\u003e and \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({\\\\varphi (X}_{2})\\\\)\\u003c/span\\u003e\\u003c/span\\u003e respectively, \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(N\\\\)\\u003c/span\\u003e\\u003c/span\\u003e is the length of the tensor \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({a}_{d}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e and \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({w}_{i}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e and \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({b}_{i}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e are the weights and biases that needs to be learned during backpropagation.\\u003c/p\\u003e \\u003c/div\\u003e \\u003cdiv id=\\\"Sec8\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003eSupport Set\\u003c/h2\\u003e \\u003cp\\u003eA support set that is necessary for few-shot validation and testing containing eight images is formed from the images in the test set. To represent each class in the support set, one image is randomly sampled from the images of the corresponding class for \\u003cem\\u003eN\\u003c/em\\u003e-way \\u003cem\\u003ek\\u003c/em\\u003e-shot validation and testing as detailed below. Whenever \\u003cem\\u003ek\\u003c/em\\u003e is greater than one, we used entirely new support set for each shot.\\u003c/p\\u003e \\u003cp\\u003e \\u003cspan type=\\\"BoldItalic\\\" class=\\\"BoldItalic\\\" name=\\\"Emphasis\\\"\\u003eN\\u003c/span\\u003e \\u003cb\\u003e-way\\u003c/b\\u003e \\u003cspan type=\\\"BoldItalic\\\" class=\\\"BoldItalic\\\" name=\\\"Emphasis\\\"\\u003ek\\u003c/span\\u003e\\u003cb\\u003e-Shot Validation and Testing\\u003c/b\\u003e\\u003c/p\\u003e \\u003cp\\u003eThe value of \\u003cem\\u003eN\\u003c/em\\u003e in \\u003cem\\u003eN\\u003c/em\\u003e-way is the number of classes which is set to eight in the present study and \\u003cem\\u003ek\\u003c/em\\u003e in \\u003cem\\u003ek\\u003c/em\\u003e-shot is between one to five for few-shot learning. If \\u003cem\\u003ek\\u003c/em\\u003e is equal to one, it is called one-shot learning and if \\u003cem\\u003ek\\u003c/em\\u003e is equal to two, it is two-shot learning, and so on. In this study, we performed 8-way 1-shot, 8-way 2-shot, 8-way 3-shot, and 8-way 5-shot validation and testing for multiclass classification of eight cell types. Finally, the class of the query image is decided based on the highest similarity with respect to the images in the support set. Mathematically, the class prediction for a query image \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({x}_{q}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e for 8-way \\u003cem\\u003ek\\u003c/em\\u003e-shot learning is given in Eq.\\u0026nbsp;(5).\\u003cdiv id=\\\"Eque\\\" class=\\\"Equation\\\"\\u003e\\u003cdiv format=\\\"TEX\\\" class=\\\"mathdisplay\\\" id=\\\"FileID_Eque\\\" name=\\\"EquationSource\\\"\\u003e\\n$${Y}_{q} = argmin\\\\left\\\\{\\\\sum _{i=1}^{k}{\\\\Psi }{({X}_{q}, {X}_{s})}_{i}\\\\right\\\\} \\\\left(5\\\\right)$$\\u003c/div\\u003e\\u003c/div\\u003e\\u003c/p\\u003e \\u003cp\\u003eAbove, \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({X}_{q}=\\\\left\\\\{{x}_{q}^{c}={x}_{q}:1\\\\le c\\\\le N \\\\right\\\\}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e, that means \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({X}_{q}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e is the set of images where the query image \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({x}_{q}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e is repeated \\u003cem\\u003eN\\u003c/em\\u003e times to match with the number of images in the support set \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({X}_{s}=\\\\left\\\\{{x}_{s}^{c}:1\\\\le c\\\\le N\\\\right\\\\}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e so that the comparison of \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({x}_{q}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e with \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({X}_{s}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e would happen in one epoch, \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({\\\\Psi }\\\\left({X}_{q}, {X}_{s}\\\\right)\\\\)\\u003c/span\\u003e\\u003c/span\\u003e is the prediction result of STN model which is a vector of \\u003cem\\u003eN\\u003c/em\\u003e similarity values between \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({X}_{q}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e and \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({X}_{s}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e. Finally, \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({Y}_{q}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e is the predicted label for the query image \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({x}_{q}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e that is between zero and seven. The calculation of overall accuracy for predicting a class \\u003cem\\u003ec\\u003c/em\\u003e during \\u003cem\\u003ek\\u003c/em\\u003e-shot validation/testing is implemented using expression (6), where \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({N}_{c}^{{\\\\prime }}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e is correctly predicted images for calss \\u003cem\\u003ec\\u003c/em\\u003e and \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({N}_{c}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e is the total number of images in class \\u003cem\\u003ec\\u003c/em\\u003e.\\u003cdiv id=\\\"Equf\\\" class=\\\"Equation\\\"\\u003e\\u003cdiv format=\\\"TEX\\\" class=\\\"mathdisplay\\\" id=\\\"FileID_Equf\\\" name=\\\"EquationSource\\\"\\u003e\\n$$Overall Accuracy = \\\\frac{\\\\sum _{i=1}^{k}({\\\\frac{{N}_{c}^{{\\\\prime }}}{{N}_{c}})}_{i}}{k} \\\\left(6\\\\right)$$\\u003c/div\\u003e\\u003c/div\\u003e\\u003c/p\\u003e \\u003c/div\\u003e \\u003cdiv id=\\\"Sec9\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003eCreation of Visual Saliency Maps\\u003c/h2\\u003e \\u003cp\\u003eFor the creation of saliency maps, to infer explainability to the network while making decisions, the output of the \\u003cem\\u003elambda\\u003c/em\\u003e layer of the STN is used as the weight tensor. The activation maps \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({A}^{H\\\\times W\\\\times C}\\\\)\\u003c/span\\u003e\\u003c/span\\u003eof the final convolution layer of the EfficientNet-B3 for the query image are multiplied with the weight vector to get weighted activation maps \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({A}_{w}^{H\\\\times W\\\\times C}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e as described by Eq.\\u0026nbsp;(7).\\u003cdiv id=\\\"Equg\\\" class=\\\"Equation\\\"\\u003e\\u003cdiv format=\\\"TEX\\\" class=\\\"mathdisplay\\\" id=\\\"FileID_Equg\\\" name=\\\"EquationSource\\\"\\u003e\\n$${A}_{w}^{H\\\\times W\\\\times i}= {A}^{H\\\\times W\\\\times i}* {a}_{d}^{i} \\\\forall i \\\\left(7\\\\right)$$\\u003c/div\\u003e\\u003c/div\\u003e\\u003c/p\\u003e \\u003cp\\u003eIn Eq.\\u0026nbsp;(5), \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({a}_{d}^{i}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e is the weight tensor as already described in Eq.\\u0026nbsp;(3). Afterward, an average activation map \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({A}_{m}^{H\\\\times W}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e of spatial size of \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(H\\\\times W\\\\)\\u003c/span\\u003e\\u003c/span\\u003e is obtained by averaging all \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(C\\\\)\\u003c/span\\u003e\\u003c/span\\u003e weighted activation maps as described in Eq.\\u0026nbsp;(8).\\u003cdiv id=\\\"Equh\\\" class=\\\"Equation\\\"\\u003e\\u003cdiv format=\\\"TEX\\\" class=\\\"mathdisplay\\\" id=\\\"FileID_Equh\\\" name=\\\"EquationSource\\\"\\u003e\\n$${A}_{m}^{H\\\\times W}=ReLU\\\\left(\\\\frac{1}{C}\\\\sum _{i=1}^{C}{A}_{w}^{H\\\\times W\\\\times i} \\\\right) \\\\left(8\\\\right)$$\\u003c/div\\u003e\\u003c/div\\u003e\\u003c/p\\u003e \\u003cp\\u003eThe negative values in the mean activation map are removed using the \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(ReLU\\\\)\\u003c/span\\u003e\\u003c/span\\u003e (rectified linear unit) activation function, which is given in Eq.\\u0026nbsp;(9).\\u003cdiv id=\\\"Equi\\\" class=\\\"Equation\\\"\\u003e\\u003cdiv format=\\\"TEX\\\" class=\\\"mathdisplay\\\" id=\\\"FileID_Equi\\\" name=\\\"EquationSource\\\"\\u003e\\n$$ReLU\\\\left(z\\\\right)= \\\\left\\\\{\\\\begin{array}{c}z if z\\u0026gt;0\\\\\\\\ 0 if z\\\\le 0\\\\end{array}\\\\right. \\\\left(9\\\\right)$$\\u003c/div\\u003e\\u003c/div\\u003e\\u003c/p\\u003e \\u003cp\\u003eFor EfficientNet-B3, \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(H\\\\times W\\\\times C\\\\)\\u003c/span\\u003e\\u003c/span\\u003e = \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(10\\\\times 10\\\\times 1536\\\\)\\u003c/span\\u003e\\u003c/span\\u003e. The depth of the activation maps \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(C\\\\)\\u003c/span\\u003e\\u003c/span\\u003e and length of the weight tensor \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(N\\\\)\\u003c/span\\u003e\\u003c/span\\u003e are identical. Finally, the 10\\u0026times;10 coarse activation map \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({A}_{m}^{H\\\\times W}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e is resized to match with the spatial resolution of the input query image which is 300\\u0026times;300 using python based \\u003cem\\u003escikit-image\\u003c/em\\u003e toolbox. For identification of highly activated regions in the query image when it is compared with the images in the support set, the resized heatmaps are overlaid onto the corresponding RGB images in the support set showing the most similar/dissimilar regions.\\u003c/p\\u003e \\u003c/div\\u003e\"},{\"header\":\"Results\",\"content\":\"\\u003cp\\u003eThe 8-way few-shot validation with the overall accuracies of the proposed STN architecture using EfficientNet-B3 as the base model are given in Table\\u0026nbsp;\\u003cspan refid=\\\"Tab2\\\" class=\\\"InternalRef\\\"\\u003e2\\u003c/span\\u003e. For different combinations of mini batch gradient descent optimizers and other model hyperparameters where \\u003cem\\u003eAdadelta\\u003c/em\\u003e performed relatively better are ascribed. Hence, the selected hyperparameters for few-shot testing along with \\u003cem\\u003eAdadelta\\u003c/em\\u003e optimizer are mini-batch size: 16, learning rate: 1.0, number of epochs: 15. The overall validation accuracy values are in the range of 94.7%-97.1%. The 8-way few-shot test overall accuracies are given in Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig6\\\" class=\\\"InternalRef\\\"\\u003e6\\u003c/span\\u003e and the best overall test accuracy over 97.2% is obtained when \\u003cem\\u003ek\\u003c/em\\u003e is equal to 3. The few-shot test overall accuracy for the classification of white blood cells alone is 97.7%. The classification value is highest for platelets with 99.8%. Table\\u0026nbsp;\\u003cspan refid=\\\"Tab3\\\" class=\\\"InternalRef\\\"\\u003e3\\u003c/span\\u003e shows the results for overall test accuracies obtained in comparison with the previous studies. The number of images used for training, validation, and testing are also mentioned for a fair comparison.\\u003c/p\\u003e \\u003cp\\u003eThe saliency maps highlighting the most discriminative/similar regions between the query image and the images in the support sets for basophil and erythroblast cells are given in Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig3\\\" class=\\\"InternalRef\\\"\\u003e3\\u003c/span\\u003e and Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig4\\\" class=\\\"InternalRef\\\"\\u003e4\\u003c/span\\u003e respectively. The 1D tensor \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({a}_{d}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e is reshaped into a 2D tensor of size 48 \\u0026times; 32 and is also plotted along with the activation maps. Further, Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig5\\\" class=\\\"InternalRef\\\"\\u003e5\\u003c/span\\u003e shows the saliency maps highlighting the regions for each cell type comparatively with another image of the same cell type in the support set to understand the reason for their similarity. As anticipated, the regions where most of the representative features located for decision making are highlighted and those regions are cell bodies and their surroundings.\\u003c/p\\u003e \\u003cp\\u003e \\u003cdiv class=\\\"gridtable\\\"\\u003e\\u003ctable float=\\\"Yes\\\" id=\\\"Tab2\\\" border=\\\"1\\\"\\u003e \\u003ccaption language=\\\"En\\\"\\u003e \\u003cdiv class=\\\"CaptionNumber\\\"\\u003eTable 2\\u003c/div\\u003e \\u003cdiv class=\\\"CaptionContent\\\"\\u003e \\u003cp\\u003eDifferent optimizers and hyperparameters along with accuracy values during Siamese network training and overall accuracy values during 8-way 1-shot, 2-shot, 3-shot, and 5-shot validation on 1000 images with 125 samples from each class. \\u003cem\\u003eSGD\\u003c/em\\u003e: stochastic gradient descent, \\u003cem\\u003elr\\u003c/em\\u003e: learning rate, \\u003cem\\u003ene\\u003c/em\\u003e: number of epochs, \\u003cem\\u003embs\\u003c/em\\u003e: mini batch size.\\u003c/p\\u003e \\u003c/div\\u003e \\u003c/caption\\u003e \\u003ccolgroup cols=\\\"6\\\"\\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=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c4\\\" colnum=\\\"4\\\"\\u003e\\u003c/div\\u003e \\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c5\\\" colnum=\\\"5\\\"\\u003e\\u003c/div\\u003e \\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c6\\\" colnum=\\\"6\\\"\\u003e\\u003c/div\\u003e \\u003cthead\\u003e \\u003ctr\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c1\\\" morerows=\\\"1\\\" rowspan=\\\"2\\\"\\u003e \\u003cp\\u003eOptimizer \\u0026amp; Hyperparameters\\u003c/p\\u003e \\u003c/th\\u003e \\u003cth align=\\\"left\\\" colspan=\\\"5\\\" nameend=\\\"c6\\\" namest=\\\"c2\\\"\\u003e \\u003cp\\u003eAccuracy\\u003c/p\\u003e \\u003c/th\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e\\u003cb\\u003eSiamese\\u003c/b\\u003e\\u003c/p\\u003e \\u003cp\\u003e\\u003cb\\u003etraining\\u003c/b\\u003e\\u003c/p\\u003e \\u003c/th\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e\\u003cb\\u003e8-way 1-shot validation\\u003c/b\\u003e\\u003c/p\\u003e \\u003c/th\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e\\u003cb\\u003e8-way 2-shot validation\\u003c/b\\u003e\\u003c/p\\u003e \\u003c/th\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e\\u003cb\\u003e8-way 3-shot\\u003c/b\\u003e\\u003c/p\\u003e \\u003cp\\u003e\\u003cb\\u003evalidation\\u003c/b\\u003e\\u003c/p\\u003e \\u003c/th\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c6\\\"\\u003e \\u003cp\\u003e\\u003cb\\u003e8-way 5-shot\\u003c/b\\u003e\\u003c/p\\u003e \\u003cp\\u003e\\u003cb\\u003evalidation\\u003c/b\\u003e\\u003c/p\\u003e \\u003c/th\\u003e \\u003c/tr\\u003e \\u003c/thead\\u003e \\u003ctbody\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003e\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(RMSprop\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e \\u003cp\\u003e\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\left\\\\{\\\\begin{array}{c}lr=0.001, ne=15, mbs=16\\\\\\\\ lr=0.001, ne=15, mbs=32\\\\\\\\ lr=0.0005, ne=20, mbs=64\\\\end{array}\\\\right.\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e\\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\left\\\\{\\\\begin{array}{c}99.80\\\\\\\\ 99.89\\\\\\\\ 99.93\\\\end{array}\\\\right.\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\left\\\\{\\\\begin{array}{c}95.31\\\\\\\\ 96.10\\\\\\\\ 94.63\\\\end{array}\\\\right.\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\left\\\\{\\\\begin{array}{c}96.40\\\\\\\\ 96.31\\\\\\\\ 95.50\\\\end{array}\\\\right.\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\left\\\\{\\\\begin{array}{c}96.40\\\\\\\\ 96.50\\\\\\\\ 95.60\\\\end{array}\\\\right.\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e \\u003cp\\u003e\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\left\\\\{\\\\begin{array}{c}96.51\\\\\\\\ 96.60\\\\\\\\ 95.60\\\\end{array}\\\\right.\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003e\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(Adam\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e \\u003cp\\u003e\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\left\\\\{\\\\begin{array}{c}lr=0.001, ne=15, mbs=16\\\\\\\\ lr=0.001, ne=15, mbs=32\\\\\\\\ lr=0.0005, ne=20, mbs=64\\\\end{array}\\\\right.\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\left\\\\{\\\\begin{array}{c}99.74\\\\\\\\ 99.79\\\\\\\\ 99.91\\\\end{array}\\\\right.\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\left\\\\{\\\\begin{array}{c}95.30\\\\\\\\ 93.10\\\\\\\\ 95.20\\\\end{array}\\\\right.\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\left\\\\{\\\\begin{array}{c}96.80\\\\\\\\ 95.35\\\\\\\\ 96.30\\\\end{array}\\\\right.\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\left\\\\{\\\\begin{array}{c}96.82\\\\\\\\ 95.40\\\\\\\\ 96.50\\\\end{array}\\\\right.\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e \\u003cp\\u003e\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\left\\\\{\\\\begin{array}{c}97.01\\\\\\\\ 95.40\\\\\\\\ 96.60\\\\end{array}\\\\right.\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003e\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(Nadam\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e \\u003cp\\u003e\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\left\\\\{\\\\begin{array}{c}lr=0.001, ne=15, mbs=16\\\\\\\\ lr=0.01, ne=15, mbs=32\\\\\\\\ lr=0.01, ne=20, mbs=64\\\\end{array}\\\\right.\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\left\\\\{\\\\begin{array}{c}99.84\\\\\\\\ 99.81\\\\\\\\ 99.74\\\\end{array}\\\\right.\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\left\\\\{\\\\begin{array}{c}94.70\\\\\\\\ 93.00\\\\\\\\ 93.70\\\\end{array}\\\\right.\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\left\\\\{\\\\begin{array}{c}96.20\\\\\\\\ 95.70\\\\\\\\ 94.60\\\\end{array}\\\\right.\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\left\\\\{\\\\begin{array}{c}95.70\\\\\\\\ 95.40\\\\\\\\ 96.00\\\\end{array}\\\\right.\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e \\u003cp\\u003e\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\left\\\\{\\\\begin{array}{c}96.25\\\\\\\\ 95.32\\\\\\\\ 95.40\\\\end{array}\\\\right.\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003e\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(Adadelta\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e \\u003cp\\u003e\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\left\\\\{\\\\begin{array}{c}lr=1.0, ne=15, mbs=16\\\\\\\\ lr=1.0, ne=15, mbs=32\\\\\\\\ lr=0.9, ne=20, mbs=64\\\\end{array}\\\\right.\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\left\\\\{\\\\begin{array}{c}99.97\\\\\\\\ 99.94\\\\\\\\ 99.95\\\\end{array}\\\\right.\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\left\\\\{\\\\begin{array}{c}96.10\\\\\\\\ 95.80\\\\\\\\ 96.00\\\\end{array}\\\\right.\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\left\\\\{\\\\begin{array}{c}97.01\\\\\\\\ 96.90\\\\\\\\ 96.90\\\\end{array}\\\\right.\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\left\\\\{\\\\begin{array}{c}96.90\\\\\\\\ 96.50\\\\\\\\ 97.10\\\\end{array}\\\\right.\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c6\\\"\\u003e \\u003cp\\u003e\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\left\\\\{\\\\begin{array}{c}97.03\\\\\\\\ 97.00\\\\\\\\ 97.00\\\\end{array}\\\\right.\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003c/tbody\\u003e \\u003c/colgroup\\u003e \\u003c/table\\u003e\\u003c/div\\u003e \\u003c/p\\u003e \"},{\"header\":\"Discussion\",\"content\":\"\\u003cp\\u003eIn this study, we assessed multi label classification of normal peripheral blood cell images that include white blood cells, immature granulocytes, erythroblasts, and platelets using STN architecture based on contrastive learning. Extensive evaluations for the choice of the optimizer and other model hyperparameters were performed that suggested the optimal choice based on their performance during few-shot validation. In general, the performance scores of the models given in Table\\u0026nbsp;\\u003cspan refid=\\\"Tab2\\\" class=\\\"InternalRef\\\"\\u003e2\\u003c/span\\u003e and were determined to be decent irrespective of the optimizer and other hyperparameters, with \\u003cem\\u003eAdadelta\\u003c/em\\u003e optimizer as a slightly better variation. As expected, based on \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({a}_{d}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e plots from Figs.\\u0026nbsp;\\u003cspan refid=\\\"Fig3\\\" class=\\\"InternalRef\\\"\\u003e3\\u003c/span\\u003e and \\u003cspan refid=\\\"Fig4\\\" class=\\\"InternalRef\\\"\\u003e4\\u003c/span\\u003e, it can be inferred that the embeddings are closer for similar pairs and apart for dissimilar pairs. Our proposed method performed better [\\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e] or slightly worse [\\u003cspan citationid=\\\"CR9\\\" class=\\\"CitationRef\\\"\\u003e9\\u003c/span\\u003e] compared to previous studies that were tested on the same dataset. Even though the overall few-shot testing accuracy is two percent less than the test accuracies from the reported study [\\u003cspan citationid=\\\"CR8\\\" class=\\\"CitationRef\\\"\\u003e8\\u003c/span\\u003e], our methodology is completely different where we have utilized only 6% of the data for training to further propel the feasibility of CL based models in the medical domain.\\u003c/p\\u003e \\u003cp\\u003eAs also discussed in [\\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e], due to the nucleus shape and structure similarities among igs, basophils, neutrophils, and monocytes, there is a higher percent of misclassifications among these cell types. From Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig6\\\" class=\\\"InternalRef\\\"\\u003e6\\u003c/span\\u003e, we can notice that 2.27%, 2.53% of igs are wrongly classified as monocytes and neutrophils respectively. Similarly, 2.48% of basophils are misclassified as igs. The classification accuracy for platelets is 99.8% since their size and morphology are entirely different compared to other cell types. In terms of overall test accuracy for the classification of only white blood cells, our results are comparable with several studies that used deep transfer learning approaches or much complicated neural network designs [\\u003cspan additionalcitationids=\\\"CR11 CR12\\\" citationid=\\\"CR10\\\" class=\\\"CitationRef\\\"\\u003e10\\u003c/span\\u003e\\u0026ndash;\\u003cspan citationid=\\\"CR13\\\" class=\\\"CitationRef\\\"\\u003e13\\u003c/span\\u003e]. Our study achieved an accuracy of 97.7% for white blood cell classification which is better than some of the reported accuracies 96.1% and 96.9% respectively [\\u003cspan citationid=\\\"CR10\\\" class=\\\"CitationRef\\\"\\u003e10\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR11\\\" class=\\\"CitationRef\\\"\\u003e11\\u003c/span\\u003e]. Our overall accuracy is comparable with the reported average accuracy of 98.8% in a recent study that used Siamese network [\\u003cspan citationid=\\\"CR35\\\" class=\\\"CitationRef\\\"\\u003e35\\u003c/span\\u003e], however, that study was restricted to white blood cells, and it used only 430 images for testing and moreover, no visual saliency mappings is implemented for explainability.\\u003c/p\\u003e \\u003cp\\u003e \\u003cdiv class=\\\"gridtable\\\"\\u003e\\u003ctable float=\\\"Yes\\\" id=\\\"Tab3\\\" border=\\\"1\\\"\\u003e \\u003ccaption language=\\\"En\\\"\\u003e \\u003cdiv class=\\\"CaptionNumber\\\"\\u003eTable 3\\u003c/div\\u003e \\u003cdiv class=\\\"CaptionContent\\\"\\u003e \\u003cp\\u003eComparison of our overall test accuracy in percentage, number of images used for training, validation and testing with previously reported studies. N: number of images.\\u003c/p\\u003e \\u003c/div\\u003e \\u003c/caption\\u003e \\u003ccolgroup cols=\\\"5\\\"\\u003e \\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c1\\\" colnum=\\\"1\\\"\\u003e\\u003c/div\\u003e \\u003cdiv align=\\\"char\\\" char=\\\".\\\" class=\\\"colspec\\\" colname=\\\"c2\\\" colnum=\\\"2\\\"\\u003e\\u003c/div\\u003e \\u003cdiv align=\\\"char\\\" char=\\\".\\\" class=\\\"colspec\\\" colname=\\\"c3\\\" colnum=\\\"3\\\"\\u003e\\u003c/div\\u003e \\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c4\\\" colnum=\\\"4\\\"\\u003e\\u003c/div\\u003e \\u003cdiv align=\\\"char\\\" char=\\\".\\\" class=\\\"colspec\\\" colname=\\\"c5\\\" colnum=\\\"5\\\"\\u003e\\u003c/div\\u003e \\u003cthead\\u003e \\u003ctr\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003eStudy\\u003c/p\\u003e \\u003c/th\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003eTest Accuracy (%)\\u003c/p\\u003e \\u003c/th\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003eTraining (N)\\u003c/p\\u003e \\u003c/th\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003eValidation (N)\\u003c/p\\u003e \\u003c/th\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003eTesting (N)\\u003c/p\\u003e \\u003c/th\\u003e \\u003c/tr\\u003e \\u003c/thead\\u003e \\u003ctbody\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003eAcevedo et.al., 2019 [\\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e]\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e96.20\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e11077\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e4096\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e1919\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003eUcar, 2020 [\\u003cspan citationid=\\\"CR9\\\" class=\\\"CitationRef\\\"\\u003e9\\u003c/span\\u003e]\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e97.94\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e13674\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e1709\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e1709\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003eLong et.al., 2021 [\\u003cspan citationid=\\\"CR8\\\" class=\\\"CitationRef\\\"\\u003e8\\u003c/span\\u003e]\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e99.30\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e13674\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e-\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e3418\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003eOur study\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e97.21\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e1000\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e1000\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e15092\\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\\u003eGoing beyond the accuracy values for comparison of current studies with previous works, and even though we have used 20000 generated image pairs, they are obtained just using 125 images from each class and yet achieved comparable test performance metrics both while classifying all cell types and white blood cells exclusively. After extensive few-shot validation, hyperparameters set based on the \\u003cem\\u003eAdadelta\\u003c/em\\u003e optimizer performed well and in addition, it can essentially eliminate learning rate selection by fixing it to one. Further, as expected, the 8-way \\u003cem\\u003ek\\u003c/em\\u003e-shot validation and testing values are better when \\u003cem\\u003ek\\u003c/em\\u003e is greater than 2 since the greater number of images per class in the support set could lead to greater classification performance. To our knowledge, none of the previous studies listed in Table\\u0026nbsp;\\u003cspan refid=\\\"Tab3\\\" class=\\\"InternalRef\\\"\\u003e3\\u003c/span\\u003e based on traditional CNN based classification had explored the class activation mapping using either novel methods or already existing schemes such as CAM [\\u003cspan citationid=\\\"CR24\\\" class=\\\"CitationRef\\\"\\u003e24\\u003c/span\\u003e], gradCAM [\\u003cspan citationid=\\\"CR25\\\" class=\\\"CitationRef\\\"\\u003e25\\u003c/span\\u003e], and gradCAM++ [\\u003cspan citationid=\\\"CR36\\\" class=\\\"CitationRef\\\"\\u003e36\\u003c/span\\u003e]. Our proposed saliency mapping method is simple as it does not require gradients computation unlike gradCAM. Since we adapted the \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\({a}_{d}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e tensor values as the weights to be multiplied with the final convolution layer activation maps of the query image, explicit retraining of the model with the query image and the support set is not essential. Moreover, from Figures (3), (4), and (5), we could use the proposed saliency mapping method for cell localization in the image, and eventually, the framework may be adapted for coarse level cell segmentation. In the future, the proposed framework could be tested on blood cell images of disease conditions such as anemia and leukemia.\\u003c/p\\u003e \"},{\"header\":\"Conclusions\",\"content\":\"\\u003cp\\u003eWe proposed a contrastive learning based approach that relatively require fewer samples for training with finetuned EfficientNet as the base model for classification of peripheral blood cell images with the overall few-shot test accuracy of 97.21%. We employed state-of-the-art EfficientNet-B3 deep learning model and finetuned the last ten percent layers to learn more representative features of blood cells. In our 8-way few-shot experiments, the \\u003cem\\u003eAdadelta\\u003c/em\\u003e optimizer with the choice of other hyperparameters provided slightly better results with the highest accuracy being 99.8% for classifying platelets. Three-shot and 5-shot testing accuracies are superior to 1-shot and 2-shot testing. Further, we demonstrated the creation of visual saliency maps without computing the gradients to interpret the decision of the Siamese network. Hence, our procedure may facilitate as a reference for deep CL based peripheral blood cell image classification against which the future research can be compared. In conclusion, we believe that our methodology could facilitate to development of deep learning based methods for classification of peripheral blood cells with fewer data samples and with interpretability.\\u003c/p\\u003e\"},{\"header\":\"Declarations\",\"content\":\"\\u003cp\\u003e\\u003cstrong\\u003eCompeting Interests:\\u003c/strong\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003eThe authors have no competing interests to declare.\\u0026nbsp;\\u003c/p\\u003e\\n\\u003cp\\u003e\\u003cstrong\\u003eFunding:\\u003c/strong\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003eNo funding was received for conducting this study.\\u0026nbsp;\\u003c/p\\u003e\\n\\u003cp\\u003e\\u003cstrong\\u003eCompliance with Ethical Standards\\u003c/strong\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003eThis research study was conducted retrospectively using human subject data made available in open access by Hospital Clinic Barcelona. Ethical approval was not required as confirmed by the license attached with the open access data.\\u003c/p\\u003e\\n\\u003cp\\u003e\\u003cstrong\\u003eAcknowledgements\\u003c/strong\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003eWe would like to thank Hospital Clinic Barcelona for providing the dataset and SRM University\\u0026mdash;AP for providing the research infrastructure.\\u003c/p\\u003e\"},{\"header\":\"References\",\"content\":\"\\u003col\\u003e\\u003cli\\u003e\\u003cspan\\u003eCeelie H, Dinkelaar RB, van Gelder W (2007) Examination of peripheral blood films using automated microscopy; evaluation of Diffmaster Octavia and Cellavision DM96. Journal of Clinical Pathology 60:72. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://doi.org/10.1136/JCP.2005.035402\\u003c/span\\u003e\\u003cspan address=\\\"10.1136/JCP.2005.035402\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eR\\u0026uuml;mke CL (1985) Imprecision of ratio-derived differential leukocyte counts. Blood Cells 11:311\\u0026ndash;4, 315\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eAcevedo A, Alf\\u0026eacute;rez S, Merino A, et al (2019) Recognition of peripheral blood cell images using convolutional neural networks. Computer Methods and Programs in Biomedicine 180:105020. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://doi.org/10.1016/J.CMPB.2019.105020\\u003c/span\\u003e\\u003cspan address=\\\"10.1016/J.CMPB.2019.105020\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eKrizhevsky A, Sutskever I, Hinton GE (2012) ImageNet Classification with Deep Convolutional Neural Networks. Advances in Neural Information Processing Systems 25:\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eTan C, Sun F, Kong T, et al (2018) A Survey on Deep Transfer Learning. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 11141 LNCS:270\\u0026ndash;279. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://doi.org/10.1007/978-3-030-01424-7_27\\u003c/span\\u003e\\u003cspan address=\\\"10.1007/978-3-030-01424-7_27\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003ePathak Y, Shukla PK, Tiwari A, et al (2020) Deep Transfer Learning Based Classification Model for COVID-19 Disease. IRBM. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://doi.org/10.1016/J.IRBM.2020.05.003\\u003c/span\\u003e\\u003cspan address=\\\"10.1016/J.IRBM.2020.05.003\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eMinaee S, Kafieh R, Sonka M, et al (2020) Deep-COVID: Predicting COVID-19 from chest X-ray images using deep transfer learning. Medical Image Analysis 65:101794. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://doi.org/10.1016/J.MEDIA.2020.101794\\u003c/span\\u003e\\u003cspan address=\\\"10.1016/J.MEDIA.2020.101794\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eLong F, Peng JJ, Song W, et al (2021) BloodCaps: A capsule network based model for the multiclassification of human peripheral blood cells. Computer methods and programs in biomedicine 202:. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://doi.org/10.1016/J.CMPB.2021.105972\\u003c/span\\u003e\\u003cspan address=\\\"10.1016/J.CMPB.2021.105972\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eUcar F (2020) Deep learning approach to cell classification in human peripheral blood. 5th International Conference on Computer Science and Engineering, UBMK 2020 383\\u0026ndash;387. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://doi.org/10.1109/UBMK50275.2020.9219480\\u003c/span\\u003e\\u003cspan address=\\\"10.1109/UBMK50275.2020.9219480\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eBaydilli YY, Atila \\u0026Uuml; (2020) Classification of white blood cells using capsule networks. Computerized medical imaging and graphics: the official journal of the Computerized Medical Imaging Society 80:. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://doi.org/10.1016/J.COMPMEDIMAG.2020.101699\\u003c/span\\u003e\\u003cspan address=\\\"10.1016/J.COMPMEDIMAG.2020.101699\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eShahin AI, Guo Y, Amin KM, Sharawi AA (2019) White blood cells identification system based on convolutional deep neural learning networks. Computer methods and programs in biomedicine 168:69\\u0026ndash;80. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://doi.org/10.1016/J.CMPB.2017.11.015\\u003c/span\\u003e\\u003cspan address=\\\"10.1016/J.CMPB.2017.11.015\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eKutlu H, Avci E, \\u0026Ouml;zyurt F (2020) White blood cells detection and classification based on regional convolutional neural networks. Medical hypotheses 135:. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://doi.org/10.1016/J.MEHY.2019.109472\\u003c/span\\u003e\\u003cspan address=\\\"10.1016/J.MEHY.2019.109472\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eAlmezhghwi K, Serte S (2020) Improved Classification of White Blood Cells with the Generative Adversarial Network and Deep Convolutional Neural Network. Computational Intelligence and Neuroscience 2020:. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://doi.org/10.1155/2020/6490479\\u003c/span\\u003e\\u003cspan address=\\\"10.1155/2020/6490479\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eMedela A, Picon A, Saratxaga CL, et al (2019) Few shot learning in histopathological images: Reducing the need of labeled data on biological datasets. Proceedings - International Symposium on Biomedical Imaging 2019-April:1860\\u0026ndash;1864. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://doi.org/10.1109/ISBI.2019.8759182\\u003c/span\\u003e\\u003cspan address=\\\"10.1109/ISBI.2019.8759182\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003ePuch S, S\\u0026aacute;nchez I, Rowe M (2019) Few-Shot Learning with Deep Triplet Networks for Brain Imaging Modality Recognition. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 11795 LNCS:181\\u0026ndash;189. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://doi.org/10.1007/978-3-030-33391-1_21\\u003c/span\\u003e\\u003cspan address=\\\"10.1007/978-3-030-33391-1_21\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eChen X, Yao L, Zhou T, et al (2021) Momentum contrastive learning for few-shot COVID-19 diagnosis from chest CT images. Pattern Recognition 113:107826. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://doi.org/10.1016/J.PATCOG.2021.107826\\u003c/span\\u003e\\u003cspan address=\\\"10.1016/J.PATCOG.2021.107826\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eChopra S, Hadsell R, LeCun Y (2005) Learning a similarity metric discriminatively, with application to face verification. Proceedings \\u0026ndash; 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, CVPR 2005 I:539\\u0026ndash;546. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://doi.org/10.1109/CVPR.2005.202\\u003c/span\\u003e\\u003cspan address=\\\"10.1109/CVPR.2005.202\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eWeinberger KQ, Saul LK (2009) Distance Metric Learning for Large Margin Nearest Neighbor Classification. Journal of Machine Learning Research 10:207\\u0026ndash;244. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://doi.org/10.5555/1577069\\u003c/span\\u003e\\u003cspan address=\\\"10.5555/1577069\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eTummala S (2021) Deep Learning Framework using Siamese Neural Network for Diagnosis of Autism from Brain Magnetic Resonance Imaging. In: 2021 6th International Conference for Convergence in Technology (I2CT). IEEE, pp\\u0026nbsp;1\\u0026ndash;5\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eMadhu G, Lalith Bharadwaj B, Rohit B, et al (2021) Convolutional Siamese networks for one-shot malaria parasite recognition in microscopic images. Demystifying Big Data, Machine Learning, and Deep Learning for Healthcare Analytics 277\\u0026ndash;306. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://doi.org/10.1016/B978-0-12-821633-0.00007-6\\u003c/span\\u003e\\u003cspan address=\\\"10.1016/B978-0-12-821633-0.00007-6\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eRossi A, Hosseinzadeh M, Bianchini M, et al (2021) Multi-Modal Siamese Network for Diagnostically Similar Lesion Retrieval in Prostate MRI. IEEE Transactions on Medical Imaging 40:986\\u0026ndash;995. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://doi.org/10.1109/TMI.2020.3043641\\u003c/span\\u003e\\u003cspan address=\\\"10.1109/TMI.2020.3043641\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eLi MD, Chang K, Bearce B, et al (2020) Siamese neural networks for continuous disease severity evaluation and change detection in medical imaging. npj Digital Medicine 2020 3:1 3:1\\u0026ndash;9. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://doi.org/10.1038/s41746-020-0255-1\\u003c/span\\u003e\\u003cspan address=\\\"10.1038/s41746-020-0255-1\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eWang J, Fang Z, Lang N, et al (2017) A multi-resolution approach for spinal metastasis detection using deep Siamese neural networks. Computers in Biology and Medicine 84:137\\u0026ndash;146. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://doi.org/10.1016/J.COMPBIOMED.2017.03.024\\u003c/span\\u003e\\u003cspan address=\\\"10.1016/J.COMPBIOMED.2017.03.024\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eZhou B, Khosla A, Lapedriza A, et al (2015) Learning Deep Features for Discriminative Localization. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition 2016-December:2921\\u0026ndash;2929. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://doi.org/10.1109/CVPR.2016.319\\u003c/span\\u003e\\u003cspan address=\\\"10.1109/CVPR.2016.319\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eSelvaraju RR, Cogswell M, Das A, et al (2016) Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization. International Journal of Computer Vision 128:336\\u0026ndash;359. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://doi.org/10.1007/s11263-019-01228-7\\u003c/span\\u003e\\u003cspan address=\\\"10.1007/s11263-019-01228-7\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eChattopadhyay A, Sarkar A, Howlader P, Balasubramanian VN (2017) Grad-CAM++: Improved Visual Explanations for Deep Convolutional Networks. Proceedings \\u0026ndash; 2018 IEEE Winter Conference on Applications of Computer Vision, WACV 2018 2018-January:839\\u0026ndash;847. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://doi.org/10.1109/WACV.2018.00097\\u003c/span\\u003e\\u003cspan address=\\\"10.1109/WACV.2018.00097\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eYoo TK, Choi JY, Kim HK (2021) Feasibility study to improve deep learning in OCT diagnosis of rare retinal diseases with few-shot classification. Medical and Biological Engineering and Computing 59:401\\u0026ndash;415. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://doi.org/10.1007/S11517-021-02321-1/FIGURES/12\\u003c/span\\u003e\\u003cspan address=\\\"10.1007/S11517-021-02321-1/FIGURES/12\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eChen L, Chen J, Hajimirsadeghi H, Mori G (2020) Adapting Grad-CAM for Embedding Networks. Proceedings \\u0026ndash; 2020 IEEE Winter Conference on Applications of Computer Vision, WACV 2020 2783\\u0026ndash;2792. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://doi.org/10.1109/WACV45572.2020.9093461\\u003c/span\\u003e\\u003cspan address=\\\"10.1109/WACV45572.2020.9093461\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eTan M, Le Q V. (2019) EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. 36th International Conference on Machine Learning, ICML 2019 2019-June:10691\\u0026ndash;10700\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eAcevedo A, Merino A, Alf\\u0026eacute;rez S, et al (2020) A dataset of microscopic peripheral blood cell images for development of automatic recognition systems. Data in Brief 30:105474. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://doi.org/10.1016/J.DIB.2020.105474\\u003c/span\\u003e\\u003cspan address=\\\"10.1016/J.DIB.2020.105474\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eDauphin YN, De Vries H, Bengio Y (2015) RMSProp and equilibrated adaptive learning rates for non-convex optimization. Advances in Neural Information Processing Systems 2015-January:1504\\u0026ndash;1512\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eKingma DP, Ba JL (2014) Adam: A Method for Stochastic Optimization. 3rd International Conference on Learning Representations, ICLR 2015 - Conference Track Proceedings\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eDozat T (2015) Incorporating Nesterov Momentum into Adam\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eZeiler MD (2012) ADADELTA: An Adaptive Learning Rate Method\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eWang Y, Cao Y (2020) A computer-assisted human peripheral blood leukocyte image classification method based on Siamese network. Medical \\u0026amp; Biological Engineering \\u0026amp; Computing 2020 58:7 58:1575\\u0026ndash;1582. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://doi.org/10.1007/S11517-020-02180-2\\u003c/span\\u003e\\u003cspan address=\\\"10.1007/S11517-020-02180-2\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eChattopadhyay A, Sarkar A, Howlader P, Balasubramanian VN (2017) Grad-CAM++: Improved Visual Explanations for Deep Convolutional Networks. Proceedings \\u0026ndash; 2018 IEEE Winter Conference on Applications of Computer Vision, WACV 2018 2018-January:839\\u0026ndash;847. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://doi.org/10.1109/WACV.2018.00097\\u003c/span\\u003e\\u003cspan address=\\\"10.1109/WACV.2018.00097\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\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\":\"SRM University AP\",\"isAcceptedByJournal\":false,\"isAuthorSuppliedPdf\":false,\"isDeskRejected\":\"\",\"isHiddenFromSearch\":false,\"isInQc\":false,\"isInWorkflow\":false,\"isPdf\":false,\"isPdfUpToDate\":true,\"isWithdrawnOrRetracted\":false,\"journal\":{\"display\":true,\"email\":\"info@researchsquare.com\",\"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\":\"Siamese network, EfficientNet, few-shot learning, blood cell typing, explainability\",\"lastPublishedDoi\":\"10.21203/rs.3.rs-2215631/v1\",\"lastPublishedDoiUrl\":\"https://doi.org/10.21203/rs.3.rs-2215631/v1\",\"license\":{\"name\":\"CC BY 4.0\",\"url\":\"https://creativecommons.org/licenses/by/4.0/\"},\"manuscriptAbstract\":\"\\u003cp\\u003eAutomated classification of blood cells from microscopic images is an interesting research area owing to advancements of efficient neural network architectures. Here, we developed a few-shot contrastive learning model for the classification of peripheral blood cells including lymphocytes, monocytes, basophils, eosinophils, neutrophils, immature granulocytes, erythroblasts, and platelets using EfficientNet as a base model and contrastive loss as a loss function. A total of 17092 publicly accessible images acquired using the CellaVision DM96 were analyzed. From 125 images of each cell type, 20000 image pairs are generated for Siamese twin network (STN) training and another 125 images from each cell type are used for few-shot validation. Therefore, out of 17092 images, 6% were used for training, 6% for few-shot validation and rest 88% for few-shot testing. This architecture demonstrates an overall accuracy of 97.21% during 8-way 3-shot testing for the classification of all cell types with an accuracy of 97.72% for the classification of white blood cells alone. Further, we propose a novel class activation mapping scheme for the interpretability of the model decisions suitable for STN. To conclude, the proposed framework based on contrastive learning could be used for the fully automated self-exploratory classification and identification of peripheral blood cells.\\u003c/p\\u003e\",\"manuscriptTitle\":\"Few-Shot Learning using Siamese Twin Network for the Classification of Blood Cells\",\"msid\":\"\",\"msnumber\":\"\",\"nonDraftVersions\":[{\"code\":1,\"date\":\"2022-10-31 20:01:01\",\"doi\":\"10.21203/rs.3.rs-2215631/v1\",\"editorialEvents\":[{\"type\":\"communityComments\",\"content\":0}],\"status\":\"published\",\"journal\":{\"display\":true,\"email\":\"info@researchsquare.com\",\"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\":\"f6ce101d-1cb4-4ca3-b9e3-c7ba600a134e\",\"owner\":[],\"postedDate\":\"October 31st, 2022\",\"published\":true,\"recentEditorialEvents\":[],\"rejectedJournal\":[],\"revision\":\"\",\"amendment\":\"\",\"status\":\"posted\",\"subjectAreas\":[],\"tags\":[],\"updatedAt\":\"2022-10-31T20:01:01+00:00\",\"versionOfRecord\":[],\"versionCreatedAt\":\"2022-10-31 20:01:01\",\"video\":\"\",\"vorDoi\":\"\",\"vorDoiUrl\":\"\",\"workflowStages\":[]},\"version\":\"v1\",\"identity\":\"rs-2215631\",\"journalConfig\":\"researchsquare\"},\"__N_SSP\":true},\"page\":\"/article/[identity]/[[...version]]\",\"query\":{\"redirect\":\"/article/rs-2215631\",\"identity\":\"rs-2215631\",\"version\":[\"v1\"]},\"buildId\":\"_2-kVJe1T_tPrBINL-cwx\",\"isFallback\":false,\"isExperimentalCompile\":false,\"dynamicIds\":[84888],\"gssp\":true,\"scriptLoader\":[]}","source_license":"CC-BY-4.0","license_restricted":false}