20 March 2025 in IEEE Access
Recent advancements in medical object detection (MOD) have been propelled by the rapid evolution of deep learning (DL) technologies, revolutionizing medical imaging and diagnostic workflows. This survey comprehensively reviews a vast scope of studies across diverse imaging modalities, including X-Ray, CT, MRI, Ultrasound, and Histopathology. Notable improvements include the integration of You Only Look Once (YOLO)-based architectures, Vision Transformers (ViT), and hybrid attention mechanisms, which have significantly enhanced detection accuracy and efficiency. Standout models, such as YOLOv8m, Hybrid YOLO-NAS, and YOLOv4+ViT, have demonstrated exceptional performance, achieving mean average precision (mAP) scores between 98.6% and 99.5%. These advancements leverage sophisticated features like Cross-Stage Partial (CSP) networks, Spatial Pyramid Pooling (SPP), and Bi-Directional Feature Pyramid Networks (BiFPN) to improve feature extraction and detection in medical images. Despite these successes, challenges remain in adapting these models to resource-limited settings and ensuring their outputs are interpretable for clinicians. This survey aims to bridge the gap between theoretical progress and practical implementation by aligning cutting-edge technological developments with clinical demands. It provides a certain roadmap for future innovation in MOD, with the overarching goal of improving patient care through enhanced diagnostic capabilities.
13 March 2023 in International Journal of Innovation in Engineering
In March 2020, the world health organization introduced a new infectious pandemic called “novel coronavirus disease” or “Covid-19”, origin dates back to World War II (1939) and spread from the city of Wuhan in China (2019). The severity of the outbreak affected the health of abundant folk worldwide. This bred the emergence of unimodal artificial intelligence approaches in the diagnosis of coronavirus disease but solely led to a significant percentage of false-negative results. In this paper, we combined 2500 Covid-19 multimodal data based on Early Fusion Type-I (EFT1) architecture as a severity recognition model for the classification task. We designed and implemented one-step systems of automatic comparative machine learning (AutoCML) and automatic comparative machine learning based on important feature selection (AutoIFSCML). We utilized our posed assessment method called “Descended Composite Scores Average (DCSA)”. In AutoCML, Extreme Gradient Boost (DCSA=0.998) and in AutoIFSCML, Random Forest (DCSA=0.960) demonstrated the best performance for multimodality Covid-19 severity recognition while 70% of the characteristics with high DCSA were chosen by the internal important features selection system (AutoIFS) to enter the AutoCML system. The DCSA-based designed systems can be useful in implementing fine-tuned machine learning models in medical processes by leveraging the capacities and performances of the model in all methods. As well as, ensemble learning made sounds good among evaluated traditional models in systems.
29 October 2023 in Frontiers in Health Informatics
Introduction: Accurate diagnosis is crucial for brain tumors, given their low survival rates and high treatment costs. However, traditional methods relying on manual interpretation of medical images are time-consuming and prone to errors. Attention-based deep learning, utilizing deep neural networks to selectively focus on relevant features, offers a promising solution.Material and Methods: This paper presents an overview of recent advancements in attention-based deep learning for brain tumor image analysis. While the reviewed models have demonstrated respectable performance across different datasets, they have yet to achieve state-of-the-art results.Results: Advanced techniques, including super-resolution image reconstruction, multi-swin-transformer blocks, and spatial group-wise enhanced attention blocks, have shown improved segmentation network performance. Integration of graph attention, swin-transformer, and gradient awareness minimization with positional attention convolution blocks, self-attention blocks, and intermittent fully connected layers has considerably enhanced the efficiency of classification networks.Conclusion: While attention-based deep learning has shown improvements in performance, challenges persist. These challenges include the requirement for large datasets, resource limitations, accurate segmentation of irregularly shaped tumors, and high computational demands. Future studies should address these challenges to further enhance the efficiency of brain tumor diagnoses in real-world settings.
17 November 2026 in medRxiv
Multiplex immunohistochemistry (mIHC) enables simultaneous biomarker profiling, yet unmixing co-localized targets with identical spectra, such as dual-DAB marking of Tau and Amyloid-beta in Alzheimer's pathology, remains mathematically ill-posed. We propose MorphoStain, an unsupervised, physics-guided framework for morphological stain decomposition with epistemic uncertainty quantification. MorphoStain models stain separation via a conditional Latent Diffusion Teacher operating on a Beer-Lambert optical-density manifold, guided by pathology foundation models. To enable real-time clinical throughput, the teacher is distilled into a FastStudentU-Net (0.12M parameters, > 1,400x compression), while a Symmetric Morphological Routing Module dynamically disentangles identical chromogen overlap. Evaluated on 9,908 held-out patches from the SEA-AD cohort, MorphoStain achieves 40.56 dB PSNR and 0.9961 SSIM at 140.1 FPS, showing statistically significant gains over InverseAE (Cohen's d = 3.52) and NMF (Cohen's d = 7.02), with p < 0.001 for both comparisons.