Accurate polyp segmentation during colonoscopy is critical for the early detection of colorectal cancer and still remains challenging due to significant size, shape, and color variations, and the camouflaged nature of polyps. While lightweight baseline models such as U-Net, U-Net++, and PraNet offer advantages in terms of easy deployment and low computational cost, they struggle to deal with the above issues, leading to limited segmentation performance. In contrast, large-scale vision foundation models such as SAM, DINOv2, OneFormer, and Mask2Former have exhibited impressive generalization performance across natural image domains. However, their direct transfer to medical imaging tasks (e.g., colonoscopic polyp segmentation) is not straightforward, primarily due to the scarcity of large-scale datasets and lack of domain-specific knowledge. To bridge this gap, we propose a novel distillation framework, PolypDiFoM, that transfers the rich representations of foundation models into lightweight segmentation baselines, allowing efficient and accurate deployment in clinical settings. In particular, we infuse semantic priors from the foundation models into canonical architectures such as U-Net and U-Net++ and further perform frequency domain encoding for enhanced distillation, corroborating their generalization capability. Extensive experiments are performed across five benchmark datasets, such as Kvasir-SEG, CVC-ClinicDB, ETIS, ColonDB, and CVC-300. Notably, PolypDiFoM consistently outperforms respective baseline models significantly, as well as the state-of-the-art model, with nearly 9× reduced computation overhead. The code is available at GitHub repository.

Polyp segmentation in colonoscopy is crucial for detecting colorectal cancer. However, it is challenging due to variations in the structure, color, and size of polyps, as well as the lack of clear boundaries with surrounding tissues. Traditional segmentation models based on Convolutional Neural Networks (CNNs) struggle to capture detailed patterns and global context, limiting their performance. Vision Transformer (ViT)-based models address some of these issues but have difficulties in capturing local context and lack strong zero-shot generalization. To this end, we propose the Mamba-guided Segment Anything Model (SAM-Mamba††Code, Modes:https://github.com/TapasKumarDuttal/SAM_Mamba_2025) for efficient polyp segmentation. Our approach introduces a Mamba-Prior module in the encoder to bridge the gap between the general pre-trained representation of SAM and polyp-relevant trivial clues. It injects salient cues of polyp images into the SAM image encoder as a domain prior while capturing global dependencies at various scales, leading to more accurate segmentation results. Extensive experiments on five benchmark datasets show that SAM-Mamba outper-forms traditional CNN, ViT, and Adapter-based models in both quantitative and qualitative measures. Additionally, SAM-Mamba demonstrates excellent adaptability to unseen datasets, making it highly suitable for real-time clinical use.

Classification of glaucoma stages remains challenging due to substantial inter-stage similarities, the presence of irrelevant features, and subtle lesion size, shape, and color variations in fundus images. For this purpose, few efforts have recently been made using traditional machine learning and deep learning models, specifically convolutional neural networks (CNN). While the conventional CNN models capture local contextual features within fixed receptive fields, they fail to exploit global contextual dependencies. Transformers, on the other hand, are capable of modeling global contextual information. However, they lack the ability to capture local contexts and merely focus on performing attention in the spatial domain, ignoring feature analysis in the frequency domain. To address these issues, we present a novel dual-domain global transformer network, Glaucoformer, to effectively classify glaucoma stages. Specifically, we propose a dual-domain global transformer layer (DGTL) consisting of dual-domain channel attention (DCA) and dual-domain spatial attention (DSA) with Fourier domain feature analyzer (FDFA) as the core component and integrated with a backbone. This helps in exploiting local and global contextual feature dependencies in both spatial and frequency domains, thereby learning prominent and discriminant feature representations. A shared key-query scheme is introduced to learn complementary features while reducing the parameters. In addition, the DGTL leverages the benefits of a deformable convolution to enable the model to handle complex lesion irregularities. We evaluate our method on a benchmark dataset, and the experimental results and extensive comparisons with existing CNN and vision transformer-based approaches indicate its effectiveness for glaucoma stage classification. Also, the results on an unseen dataset demonstrate the generalizability of the model.

Accurate polyp segmentation is paramount for diagnosing colorectal cancer (CRC). Though there has been a significant stride in developing polyp segmentation methods with the help of deep learning models, the diverse changes in the shape and size of polyps during various stages still make the task more challenging. To this end, we propose a novel lightweight multiscale convolutional transformer network named MCT-Net by integrating the benefits of both convolution and transformer for accurate polyp segmentation. The MCT-Net is a U-shaped architecture, mainly comprising a multiscale encoder and a transformer decoder. The encoder facilitates learning feature representations at multiple scales and subsequently introduces a cascaded attention block to learn to emphasize only polyp regions. On the other hand, the transformer decoder fully models the long-range contextual dependencies through a modified self-attention mechanism and preserves the fine-grained contextual details through a skip connection. The MCT-Net effectively mitigates the issue faced by individual convolution operations and transformers. Quantitative and qualitative results and comparisons on three benchmark datasets confirm the effectiveness of the MCT-Net over state-of-the-art segmentation methods. Further, the ablation studies verify the impact of each introduced component in the encoder and decoder block of MCT-Net.

Classification of multiclass breast cancer through histopathological images is indispensable and poses daunting challenges due to color inconsistencies, high appearance variations, and large inter-class similarities. Regardless of the success of traditional convolutional neural networks (CNNs) and vision transformers (ViTs), they often fail to capture intricate patterns and local contextual information, respectively, while demanding high computational resources and data requirements. To mitigate these issues, this paper proposes a lightweight convo-lutional transformer network named LiCT-Net for multiclass breast cancer classification. The LiCT-Net introduces a local-global spatially-aware transformer layer and integrates it with a pre-trained FastViT model to effectively capture global and local contextual features, thereby facilitating learning fine-grained and intricate lesion patterns from histopathological images. The LiCT-Net is validated on a benchmark dataset and the experimental results and comparative analysis demonstrate its effectiveness. In specific, it achieves a higher accuracy of 96.16 %, 95.62 %, 95.25 %, and 94.21 % on$40\times, 100\times, 200\times$, and$400\times$magniflcations respectively,

Glaucoma is a progressive eye disorder that can lead to permanent vision loss if not identified and treated promptly. Thus, timely glaucoma detection is paramount to developing a more efficient treatment plan and saving vision loss. Despite the promising performance achieved by deep learning methods in specific convolutional neural networks (CNNs) for glaucoma screening using fundus images, they are confined to binary classification tasks (i.e., healthy versus glaucoma) and cannot detect glaucoma stages. However, it is challenging to diagnose the glaucoma stages accurately due to considerable interstage similarities, the subtle changes in the size of lesions, and the presence of irrelevant features. Moreover, fundus images encompass significant uncertain information, which cannot be effectively captured through conventional CNNs. To solve these problems, we present a novel fuzzy joint attention-guided network called FJA-Net for the screening of glaucoma stages. Specifically, we introduce a fuzzy joint attention module (FJAM) on top of a backbone, composed of a local–global channel and spatial attention block, to learn comprehensive feature correlations along the relevant channels and spatial positions, each followed by a fuzzy layer to reduce the uncertainty in the feature representations. The FJAM aids in learning stage-specific and fine-grained features from critical regions of the fundus images. In addition, we propose a combined loss function to train the parameters of our FJA-Net to ensure better generalization and robustness. We evaluate the proposed model on two datasets, and the results of the comparative analysis demonstrate that our FJA-Net outperforms state-of-the-art CNN-based glaucoma classification approaches.

Glaucoma is a common eye disease that leads to irreversible blindness unless timely detected. Hence, glaucoma detection at an early stage is of utmost importance for a better treatment plan and ultimately saving the vision. The recent literature has shown the prominence of CNN-based methods to detect glaucoma from retinal fundus images. However, such methods mainly focus on solving binary classification tasks and have not been thoroughly explored for the detection of different glaucoma stages, which is relatively challenging due to minute lesion size variations and high inter-class similarities. This paper proposes a global self-attention based network called GS-Net for efficient multi-stage glaucoma classification. We introduce a global self-attention module (GSAM) consisting of two parallel attention modules, a channel attention module (CAM) and a spatial attention module (SAM), to learn global feature dependencies across channel and spatial dimensions. The GSAM encourages extracting more discriminative and class-specific features from the fundus images. The experimental results on a publicly available dataset demonstrate that our GS-Net outperforms state-of- the-art methods. Also, the GSAM achieves competitive performance against popular attention modules.

Abstract Due to inter‐class homogeneity and intra‐class variability, the classification of skin lesions in dermoscopy images has remained difficult. Although deep convolutional neural networks (DCNNs) have achieved satisfactory performance for binary skin cancer classification, multiclass skin lesion classification is still an open problem due to the limited training samples and class imbalance issues. To tackle these issues, in this article, we propose a multi‐CNN ensemble approach dubbed for multiclass skin lesion classification. The includes three individual CNN models, each helping in extracting different high‐level features from skin images and thereby yielding different prediction results. First, we design a lightweight CNN model to extract prominent features and train it from scratch, which primarily aims at avoiding the data scarcity problem. Then, we ensemble two different pre‐trained CNN models with the lightweight model to improve the performance and generalization capability. The proposed ensemble approach can effectively fuse the predictions of each individual CNN model using the averaging method. The approach is validated using a benchmark data set, HAM10000, which contains skin lesion images of seven different classes. The results demonstrate that the outperforms base CNN models and state‐of‐the‐art approaches without using any external data.

Brain tumor is the most common type of cancer that causes a high mortality rate among individuals of all age groups. Hence, accurate diagnosis of brain tumor and its type at early stages is of utmost importance to preclude its severity and ultimately helps patients for timely and better treatment. The current literature has witnessed the usage of convolutional neural networks (CNN) for brain tumor classification in MR images; however, such traditional CNN methods may fail to identify the minute variations of tumor lesions. This paper proposes an attention module called channel split dual attention (CSDA) coupled with a backbone network to handle this issue. The CSDA explores more detailed and discriminative features using channel splitting and two parallel attention blocks: position attention block (PAB) and channel attention block (CAB). The PAB and CAB are introduced to capture feature dependencies in spatial and channel dimensions. Extensive experiments on a publicly available dataset show that our CDANet significantly improves the state-of-the-art CNN results and obtains higher classification accuracy than existing brain tumor detection methods. The codes and models are available at https://github.com/TapasKumarDutta1/CDANet.

Accurate and rapid diagnosis of COVID-19 is crucial for curbing its fast spread across the globe, with constant mutations leading to newer variants. Recent studies have exhibited that chest CT scans manifest clear radiological findings for the COVID-19 infected patients. Convolutional neural networks (CNN) have been used considerably for COVID-19 diagnosis; however, most CNN architectures demand a huge amount of parameters, resulting in overfitting on limited training data and a slower inference. Further, residual and densely connected neural networks such as ResNet and DenseNet have been proven to strengthen feature extraction and feature propagation but fail to fully discover both local and global representations. Moreover, few linearly stacked networks fall short in capturing and preserving multiscaled features from various receptive fields. This paper proposes a new CNN architecture called global dense multiscale feature learning network (GDenseMNet) for COVID-19 detection from CT images that effectively incorporates global dense connections while capturing multiscaled features. The GDenseMNet model comprises multiscale local feature extraction (MLF) blocks that capture local features of various size receptive fields using multiple filters and residual skip connections. The global dense connections between these blocks further enable global feature learning capability. The proposed architecture is lightweight, end-to-end learnable, and validated using the SARS-CoV-2 CT-Scan dataset. Experimental results demonstrate that the GDenseMNet model achieves promising detection performance compared to state-of-the-art CNN approaches and hence, it can be utilized as an effective tool real-time COVID-19 diagnosis.

Timely and accurate diagnosis of coronavirus disease 2019 (COVID-19) is crucial in curbing its spread. Slow testing results of reverse transcription-polymerase chain reaction (RT-PCR) and a shortage of test kits have led to consider chest computed tomography (CT) as an alternative screening and diagnostic tool. Many deep learning methods, especially convolutional neural networks (CNNs), have been developed to detect COVID-19 cases from chest CT scans. Most of these models demand a vast number of parameters which often suffer from overfitting in the presence of limited training data. Moreover, the linearly stacked single-branched architecture based models hamper the extraction of multi-scale features, reducing the detection performance. In this paper, to handle these issues, we propose an extremely lightweight CNN with multi-scale feature learning blocks called as MFL-Net. The MFL-Net comprises a sequence of MFL blocks that combines multiple convolutional layers with 3 ×3 filters and residual connections effectively, thereby extracting multi-scale features at different levels and preserving them throughout the block. The model has only 0.78M parameters and requires low computational cost and memory space compared to many ImageNet pretrained CNN architectures. Comprehensive experiments are carried out using two publicly available COVID-19 CT imaging datasets. The results demonstrate that the proposed model achieves higher performance than pretrained CNN models and state-of-the-art methods on both datasets with limited training data despite having an extremely lightweight architecture. The proposed method proves to be an effective aid for the healthcare system in the accurate and timely diagnosis of COVID-19.

Abstract Facial expression recognition has been a long‐standing problem in the field of computer vision. This paper proposes a new simple scheme for effective recognition of facial expressions based on a hybrid feature descriptor and an improved classifier. Inspired by the success of stationary wavelet transform in many computer vision tasks, stationary wavelet transform is first employed on the pre‐processed face image. The pyramid of histograms of orientation gradient features is then computed from the low‐frequency stationary wavelet transform coefficients to capture more prominent details from facial images. The key idea of this hybrid feature descriptor is to exploit both spatial and frequency domain features which at the same time are robust against illumination and noise. The relevant features are subsequently determined using linear discriminant analysis. A new least squares support vector machine parameter tuning strategy is proposed using a contemporary optimisation technique called Jaya optimisation for classification of facial expressions. Experimental evaluations are performed on Japanese female facial expression and the Extended Cohn–Kanade (CK+) datasets, and the results based on 5‐fold stratified cross‐validation test confirm the superiority of the proposed method over state‐of‐the‐art approaches.

Segmentation of handwritten words into isolated characters and their recognition are challenging due to the presence of high variability and cursiveness in Indian scripts. The complex shapes and availability of numerous atomic character classes, compound characters, modifiers, ascendants, and descendants make the recognition task even more difficult. A holistic approach effectively tackles such issues by avoiding the character‐level segmentation and the earlier holistic methods have been mostly developed using multi‐stage machine learning architecture. In this study, a deep convolutional neural network‐based holistic method termed ‘H‐WordNet’ is proposed for handwritten word recognition. The H‐WordNet model includes merely four convolutional layers and one fully connected layer to effectively classify the word images', which lead to a significant reduction in parameters. The efficacy of different pooling operations with the proposed model is investigated. The main purpose of this study is to avoid the need for handcrafted feature extraction and obtain a more stable and generalised system for word recognition. The proposed model is evaluated using a standard handwritten Bangla word database (CMATERdb2.1.2), which contains 18000 Bangla word images of 120 different categories and it obtained a higher recognition accuracy of 96.17% when compared to recent state‐of‐the‐art methods.

Automated diagnosis of pathological brain not only reduces the diagnostic error significantly but also improves the patient's quality of life, thereby addressing the sustainability issues. The last few decades have witnessed an intensive research on binary classification of brain magnetic resonance (MR) images. Multiclass classification of pathological brain MR images is a more challenging task and the literature on this problem is still in its infancy. In this paper, we propose a new automated diagnosis system to classify the brain MR images into five different categories. Texture features within MR images play a significant role in accurate and efficient pathological brain detection. This work presents the extraction of such vital texture features by calculating the entropy over the curvelet subbands. Two faster and simpler strategies of fast curvelet transform are separately employed for feature extraction and the derived features are termed as FCEntF-I and FCEntF-II. The features are finally subjected to kernel extreme learning machine (K-ELM) for classification. The effectiveness of the proposed scheme is evaluated on multiclass as well as binary brain MR datasets. Comparisons with state-of-the-art methods indicate the superiority of the proposed scheme. The discriminatory potential of FCEntF-I and FCEntF-II features is found better than its counterparts.

This paper aims at developing a new method for automated diagnosis of pathological brain using magnetic resonance imaging (MRI). The method derives features using unequally-spaced FFT based fast discrete curvelet transform (FDCT- USFFT). Thereafter, a reduced feature set is obtained using PCA+LDA algorithm. Finally, for classification, we hybridize regularized extreme learning machine and sine cosine algorithm (SCA-RELM) which aims at overcoming the drawbacks of conventional ELM and other classical learning algorithms. We evaluate our proposed scheme on three well-studied datasets and observe that it earns significant improvements over the existing methods. Moreover, the effectiveness of proposed SCA-RELM paradigm is tested against other learning algorithms for single layer feed-forward neural network. Our system will aid the clinicians to effectively diagnose pathological brain.