
In cancer genomics, the relationship between gene expression and cancer development remains complex. Identifying key candidate driver genes is essential for advancing targeted therapeutic strategies. Although microarray technology provides valuable insights into gene expression patterns, its high-dimensional nature poses significant challenges for reliable gene selection. To address this, we propose CGI_MMLP, a unified framework for cancer gene prioritization in high-dimensional, low-sample-size (HDLSS) settings. Rather than relying on a single algorithm, the proposed framework integrates metaheuristic-guided gene selection with machine learning within a wrapper-based strategy, enabling systematic exploration of the feature space and improved classification performance. To bridge computational prediction with biological relevance assessment, we further introduce a structured pathway-enrichment-based validation framework that systematically prioritizes genes, identifies significantly enriched cancer-related pathways, integrates information from multiple biological knowledge bases, and interprets gene-pathway relationships to enhance biological interpretability A weighted ranking mechanism is introduced to prioritize genes based on both classification performance and selection stability across multiple metaheuristic-classifier combinations. The top-ranked genes are subsequently evaluated through the proposed biological relevance assessment workflow to establish their associations with cancer-related pathways and biological processes. The proposed framework is evaluated on three DNA microarray HDLSS gene expression cancer datasets. Experimental results demonstrate that it achieves competitive classification accuracy compared with state-of-the-art methods while consistently identifying biologically relevant genes associated with significantly enriched cancer-related pathways, supporting robust cancer gene prioritization.
Deep neural networks have achieved remarkable performance in computer vision tasks. However, the high cost of annotations in medical domains motivates interest in semi-supervised methods which leverage limited annotations along side large unlabeled data. In this paper, we develop a semi-supervised model for pathology detection in wireless capsule endoscopy based on the Teacher Student architecture. Our method combines tailored augmentations with local consistency-based embedding structuring to improve detection under limited supervision. The proposed method aims to improve detection performance and pseudo-label reliability through class-aware, region-level embedding consistency in the embedding space, where teacher-generated pseudo-labeled pathology regions are aligned with ground-truth regions from labeled instances of the same class. Multiple labeled instances are used to construct a more representative class-specific reference, providing an additional semantic constraint beyond conventional detection consistency. We evaluated our method using three losses to maximize embedding similarities based on cosine similarity, contrastive InfoNCE and VICREG. Our method generalizes across three medical domains: endoscopy (Kvasir-capsule and SEE-AI datasets), radiography (VinDr-SpineXR), and microscopy (Parasites). On the endoscopic datasets Kvasir-Capsule and SEE-AI, we achieve gains over Unbiased Teacher ranging from +3.54% to +6.01% across multiple metrics, including AP, AP50, AP75, APm, and APl. We achieve competitive performance with fully supervised methods YOLOX-1 and YOLOv5 on the VinDr-SpineXR dataset. Furthermore, we achieve gains in mAP of +6.27% and +6.70% on the parasite dataset over semi-supervised consistent Teacher and Soft-Teacher, respectively. Our code is available at https://github.com/agossouema2011/SSL_Emb_Similarity
High-resolution histopathology images are essential for visualizing tissue architecture and cellular morphology; however, their acquisition increases scanning time, storage, and data transfer requirements. Super-resolution (SR) offers a computational alternative; however, conventional SR networks are generally optimized for reconstruction or perceptual quality without explicitly targeting the preservation of histopathological morphology. This study proposes a histology-aware hybrid CNN–SwinIR framework for ×2 and ×4 histopathological image super-resolution, integrating morphology-aware processing at multiple stages. The architecture combines a CNN-based morphology stem for capturing local textures and edge-sensitive features, a lightweight SwinIR-inspired module for modeling a broader tissue context, and a histology-aware residual detail adapter that refines morphology-relevant information through local texture features, fixed edge guidance, and attention-based refinement before PixelShuffle reconstruction. The framework was evaluated on breast PR-IHC and lung and colon histopathological images under synthetic degradation, cross-tissue transfer, and real cross-scanner conditions. Under synthetic degradation, the proposed model achieved PSNR/SSIM values of 36.98 dB/0.91 and 34.14 dB/0.87 on PR-IHC images at ×2 and ×4, respectively, and 37.89 dB/0.94 and 34.40 dB/0.90 on lung images, respectively. It also generalized effectively to unseen colon tissue, reaching 35.21 dB PSNR and 0.93 SSIM at ×4. Beyond conventional image quality metrics, downstream segmentation with pathologist-verified masks showed that the reconstructed images retained morphology relevant to analysis, achieving a Dice score of 0.8683 compared with 0.8721 for the original HR images, with comparable HD95 and component-count errors. Performance decreased on genuinely acquired lower-magnification images from a different scanning setting, indicating a remaining synthetic-to-real and cross-scanner gap. Overall, the proposed architecture advances histopathological SR by combining local morphology extraction, contextual transformer modeling, and explicit structure-aware detail refinement within a unified, compact, and parameter-efficient reconstruction framework, while demonstrating the preservation of downstream morphological information across synthetic, cross-tissue, and real acquisition settings.
In this study, a heterogeneous module optimization approach for lung image segmentation is introduced, where different optimizers along with their corresponding hyperparameters are chosen for the modules separately. Instead of tuning one optimizer to optimize the whole network, the encoder, bottleneck, and decoder modules of the network can have their own optimizers chosen among Adam, RAdam, AdamW, Lion, and Lookahead. In order to select the most effective optimizer and its related hyperparameters, including the learning rate, weight decay, β1, β2, and Lookahead parameters (k, α) if needed, PSO and DE algorithms are utilized. The effectiveness of our approach is shown through FF-UNet model using both Montgomery and Chest X-ray datasets. We also use U-Net in order to see how well our framework can be applied to other network architectures. For the FF-UNet network, DSC scores of 96.09% and 95.30% are obtained for the Montgomery and Chest X-ray datasets, respectively, while the transferred configurations to U-Net result in DSC scores of 96.00% and 96.06%. Furthermore, an ablation study comparing global tuning, module-specific hyperparameter tuning, and heterogeneous optimization highlights the effectiveness of adapting the training strategy to the different network modules.
Chronic kidney disease (CKD) is a progressive and often asymptomatic disorder that is frequently diagnosed only at advanced stages, limiting treatment effectiveness and increasing the burden on healthcare systems. Consequently, there is a growing need for reliable, non-invasive, and scalable approaches for early detection. In recent years, retinal fundus imaging has emerged as a promising and cost-effective modality, as retinal microvascular alterations can reflect systemic vascular changes associated with kidney dysfunction. Although CKD and diabetic retinopathy (DR) are distinct clinical conditions, they share common microvascular pathological mechanisms, particularly in individuals with diabetes. As a result, many retinal biomarkers, publicly available datasets, image analysis techniques, and artificial intelligence (AI) models originally developed for DR provide valuable methodological foundations for CKD detection. This paper presents a comprehensive narrative review of AI-assisted CKD detection using retinal fundus images, covering the complete retinal image analysis pipeline, including image preprocessing, vessel segmentation, artery–vein separation, retinal biomarker extraction, and AI-based classification methods. While the primary focus of this review is CKD detection, relevant DR studies are also discussed because they have substantially advanced retinal image analysis methodologies, publicly available datasets, and retinal biomarker research that are directly applicable to CKD. The review critically examines publicly available datasets, key retinal microvascular biomarkers such as vessel calibre, tortuosity, fractal dimension, vessel density, and arteriovenous ratio, together with recent advances in machine learning and deep learning approaches. Unlike existing reviews that primarily focus on ocular disease analysis, this review provides a CKD-specific perspective while clarifying the role of DR research in advancing retinal imaging-based CKD detection. It further identifies current research challenges, including limited CKD-specific datasets, cross-center generalizability, vessel-specific biomarker analysis, and multimodal data integration. Overall, the findings highlight the considerable potential of AI-assisted retinal fundus imaging as a scalable, non-invasive, and clinically applicable approach for early CKD screening and future precision healthcare.