Medical imaging plays a crucial role in modern healthcare, serving as a fundamental component for diagnosis and treatment. The selection of imaging modalities for specific diagnostic tasks often involves balancing accessibility, cost, and performance. In this study, we propose a novel approach that utilizes knowledge from a standard modality characterized by higher performance but lower feasibility to enhance the effectiveness of a meta-modality, which is more accessible yet under-performing. Our focus is on applying deep learning techniques in medical image diagnosis by building a lightweight model. This lightweight mapping model utilizes potential representations leveraging the standard modality to improve the training of a model that relies solely on meta-modality. The effectiveness of our approach are illustrated through its use in a clinical setting for multi-task classification of skin lesions, relying on both clinical and dermoscopic images. Our results indicate a notable enhancement within the diagnostic accuracy of the meta-modality, attained without the need for the standard modality during inference. The balanced accuracy increased from 0.87 for clinical images alone to 0.92 for the combined model using both modalities. Furthermore, the experiments were performed repeatedly to ensure their consistency under several random weight configurations.

Deep learning has significantly improveded medical image classification through Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs). However, CNNs struggle to capture long-range dependencies, while ViTs require substantial computational demands. To address these limitations, we introduce an Adaptive Lightweight sequential space model, named ALMamba. This novel architecture integrates Structured State Space Model (SSMs) with an Adaptive Channel Modulation and Enhancement Block, within a unified SS-ACME-SSM block. By using channel splitting and separate global and local feature processing, ALMamba effectively captures both detailed and broad image features. To prevent information loss, we implement a channel shuffle operation followed by channel concatenation. The proposed ALMamba framework significantly reduces the model parameters and computational complexity, making it suitable for efficient medical imaging tasks. Extensive experiments on six diverse medical image datasets demonstrate, ALMamba competitive performance, achieving 1.93G Flops and 8.60M parameters, surpassing several state-of-the-art methods in medical image classification. This study shows a new benchmark and offers valuable insights for advancing SSM-based AI algorithms in the healthcare domain.

Accurately capturing the complex structures of human body components such as the bones, lungs, and abdominal organs is essential for diagnostic applications. Computed tomography (CT) enables detailed imaging; however, excessive radiation exposure poses health risks, whereas lower radiation levels introduce artifacts, degrading image quality. To address these challenges, we propose multipath feature fusion encoder-decoder super-resolution (MF2EDSR), a multipath encoder-decoder network designed for medical image super-resolution. The model integrates a coarse-to-fine learning strategy, allowing progressive feature refinement through a hierarchical manner. Our approach utilizes a two-path feature fusion mechanism, where a primary path focuses on large-scale structural details, and a refinement path enhances finer textures. To further improve feature representation, the feature extraction residual network is designed as a decoder. This incorporates a global enhancement module that preserves low-frequency global structures and a local enhancement module that refines high-frequency textures. Skip connections ensure stable gradient flow, mitigating the vanishing gradient issue. Extensive experiments on benchmark datasets, including medical CT, MRI, and natural image datasets, demonstrate the effectiveness of our method. The proposed MF2EDSR architecture outperforms state-of-the-art models. The results confirm that our model excels in reconstructing high-quality images with superior edge definition and minimal artifacts, making it a promising solution for medical image enhancement.

The recent emergence of the monkeypox virus has attracted significant attention due to its serious effects, including pneumonia, retinal problems, secondary skin infections, and rectal swelling, leading to discomfort and urinary issues. Therefore, accurate identification is crucial, but the lack of comprehensive databases for each condition presents a challenge. Although Generative Adversarial Networks (GANs) show promise for improving data through augmentation, their training is hindered by differences between random noise and image structures. Because GAN-based models take random noise as input for the generators, they must acquire an adaptive function to translate this noise from randomness to meaningful visuals, leads to difficulty in training and converging. To tackle this issue, a new technique is proposed for data augmentation using a Densely Connected Decoder Encoder Generative Adversarial Network (D3EGAN). This approach combines adversarial training with variational Bayesian inference to enhance image production performance. D3EGAN uses a reversed encoder–decoder decoder–encoder framework which is pre-trained to transform random noise into more informative vectors, improving the generator’s learning ability. With this enhanced dataset, advanced deep learning models are employed such as VGG-19, ResNet50, InceptionV3, and EfficientNet to distinguish monkeypox from other diseases. Our study’s results demonstrate significant improvements in image clarity and detection accuracy, surpassing current techniques’ performance. This study highlights the effectiveness of our methodology in improving timely identification and intervention for monkeypox and related diseases, addressing a significant global health issue.

The recent surge in monkeypox (mpox) cases across various nations has escalated into a significant public health challenge, underscoring the imperative for timely detection and diagnosis. In light of this, our research focused on identifying the most efficient deep learning model tailored for mpox detection. Clinically, the progression of monkeypox is characterized by four distinct stages: macular, papular, vesicular, and pustular. In this study, we explore the viability of leveraging a Deep Convolutional Generative Adversarial Network (DCGAN) in conjunction with a conditional vector to strengthen the diagnostic precision of monkeypox. By utilising the remarkable capabilities of GANs, we generate synthetic images mirroring monkeypox skin lesions, thereby amplifying our limited dataset. This approach of utilizing a conditional DCGAN, when paired with stage-specific conditions, results in a significant boost in classification accuracy, elevating it from 0.7532 to 0.8734, complemented by precision 0.91, recall 0.878 F1-score 0.8936. These promising results underscore the advantages of deploying GANs for data augmentation in biomedical image classification tasks, when compared with well known classification models. These models assures the potential of deep learning-assisted diagnostics in dermatology. Especially in therapeutic contexts, where real data may be sparse, the addition of synthetic images can act as a valuable resource, paving the way for enhanced diagnostic instruments.

Low-resolution MRI images in the through-plane direction impose constraints on both time and cost. This insufficient level of resolution does not meet the requirements for high-resolution imaging that is necessary for early diagnosis of brain diseases and morphometric analysis. Conventional single-image super-resolution (SISR) techniques have significant obstacles, such as effectively integrating intricate local details with the overall anatomical structure. In order to address these difficulties, we propose a novel technique known as the Progressively Serial-Stacked Encoder Decoder (PSSED) to improve the resolution of brain MRI scans. This approach employs convolutional blocks to extract specific information and enhance the quality of the image. Our strategy has been proven to be superior to previous Single Image Super-Resolution (SISR) methods based on extensive experimentation using publically available datasets. Our findings highlight the potential of deep learning techniques to address the constraints of low-resolution MRI images, paving the path for improved diagnostic accuracy and lower radiation exposure in clinical practice.

A brain tumour is an abnormal development of abnormal brain cells that causes harm to the blood vessels and neurons of the human body. The two forms of brain tumours are benign and malignant. It is vital to diagnose a brain tumour early in order to save the life of someone who is in danger. According to research conducted in the United Kingdom, only approximately 15 persons out of 100 will be able to survive for ten years or longer after being recognised. Because of the advent of Artificial Intelligence, Deep Learning models are now being utilised to identify brain cancers using Magnetic Resonance Imaging (MRI) images. MRI is a type of scanning method that uses high magnetic fields and radio waves to obtain detailed pictures of the interior body. The notion of Transfer Learning and Residual Networks were employed in the suggested work. The model utilised is a classification task that uses MRI images to determine whether or not a tumour is present. We may use the same transfer learning techniques to precisely pinpoint the tumour once it has been found. We utilised several optimizers in Resnet-50 to see how our model responds to changes in optimization strategies. Our suggested model has accuracy with Amsgrad optimizer compared with Adam optimizer, demonstrating improved robustness over the state-of-the-art techniques.