A hybrid explainable model based on advanced machine learning and deep learning models for classifying brain tumors using MRI images

Scientific Reports, Journal Year: 2025, Volume and Issue: 15(1)

Published: Jan. 10, 2025

Brain tumors present a significant global health challenge, and their early detection accurate classification are crucial for effective treatment strategies. This study presents novel approach combining lightweight parallel depthwise separable convolutional neural network (PDSCNN) hybrid ridge regression extreme learning machine (RRELM) accurately classifying four types of brain (glioma, meningioma, no tumor, pituitary) based on MRI images. The proposed enhances the visibility clarity tumor features in images by employing contrast-limited adaptive histogram equalization (CLAHE). A PDSCNN is then employed to extract relevant tumor-specific patterns while minimizing computational complexity. RRELM model proposed, enhancing traditional ELM improved performance. framework compared with various state-of-the-art models terms accuracy, parameters, layer sizes. achieved remarkable average precision, recall, accuracy values 99.35%, 99.30%, 99.22%, respectively, through five-fold cross-validation. PDSCNN-RRELM outperformed pseudoinverse (PELM) exhibited superior introduction led enhancements performance parameters sizes those models. Additionally, interpretability was demonstrated using Shapley Additive Explanations (SHAP), providing insights into decision-making process increasing confidence real-world diagnosis.

Language: Английский

---

D-YOLO: A Lightweight Model for Strawberry Health Detection

Agriculture, Journal Year: 2025, Volume and Issue: 15(6), P. 570 - 570

Published: March 7, 2025

In complex agricultural settings, accurately and rapidly identifying the growth health conditions of strawberries remains a formidable challenge. Therefore, this study aims to develop deep framework, Disease-YOLO (D-YOLO), based on YOLOv8s model monitor status strawberries. Key innovations include (1) replacing original backbone with MobileNetv3 optimize computational efficiency; (2) implementing Bidirectional Feature Pyramid Network for enhanced multi-scale feature fusion; (3) integrating Contextual Transformer attention modules in neck network improve lesion localization; (4) adopting weighted intersection over union loss address class imbalance. Evaluated our custom strawberry disease dataset containing 1301 annotated images across three fruit development stages five plant states, D-YOLO achieved 89.6% mAP train set 90.5% test while reducing parameters by 72.0% floating-point operations 75.1% compared baseline YOLOv8s. The framework’s balanced performance efficiency surpass conventional models including Faster R-CNN, RetinaNet, YOLOv5s, YOLOv6s, comparative trials. Cross-domain validation maize demonstrated D-YOLO’s superior generalization 94.5% mAP, outperforming YOLOv8 0.6%. (89.6% training mAP) models, YOLOv8s, This lightweight solution enables precise, real-time crop monitoring. proposed architectural improvements provide practical paradigm intelligent detection precision agriculture.

Language: Английский

---

Medivision: Empowering Colorectal Cancer Diagnosis and Tumor Localization Through Supervised Learning Classifications and Grad-CAM Visualization of Medical Colonoscopy Images

Cognitive Computation, Journal Year: 2025, Volume and Issue: 17(2)

Published: March 21, 2025

Language: Английский

---

Brain tumor segmentation using multi-scale attention U-Net with EfficientNetB4 encoder for enhanced MRI analysis

Scientific Reports, Journal Year: 2025, Volume and Issue: 15(1)

Published: March 22, 2025

Abstract Accurate brain tumor segmentation is critical for clinical diagnosis and treatment planning. This study proposes an advanced framework that combines Multiscale Attention U-Net with the EfficientNetB4 encoder to enhance performance. Unlike conventional U-Net-based architectures, proposed model leverages EfficientNetB4’s compound scaling optimize feature extraction at multiple resolutions while maintaining low computational overhead. Additionally, Multi-Scale Mechanism (utilizing $$1\times 1, 3\times 3$$ , $$5\times 5$$ kernels) enhances representation by capturing boundaries across different scales, addressing limitations of existing CNN-based methods. Our approach effectively suppresses irrelevant regions localization through attention-enhanced skip connections residual attention blocks. Extensive experiments were conducted on publicly available Figshare dataset, comparing EfficientNet variants determine optimal architecture. demonstrated superior performance, achieving Accuracy 99.79%, MCR 0.21%, Dice Coefficient 0.9339, Intersection over Union (IoU) 0.8795, outperforming other in accuracy efficiency. The training process was analyzed using key metrics, including Coefficient, dice loss, precision, recall, specificity, IoU, showing stable convergence generalization. method evaluated against state-of-the-art approaches, surpassing them all accuracy, mean IoU. demonstrates effectiveness robust efficient tumors, positioning it as a valuable tool research applications.

Language: Английский

---

Improving Malaria diagnosis through interpretable customized CNNs architectures

Scientific Reports, Journal Year: 2025, Volume and Issue: 15(1)

Published: Feb. 22, 2025

Abstract Malaria, which is spread via female Anopheles mosquitoes and brought on by the Plasmodium parasite, persists as a serious illness, especially in areas with high mosquito density. Traditional detection techniques, like examining blood samples microscope, tend to be labor-intensive, unreliable necessitate specialized individuals. To address these challenges, we employed several customized convolutional neural networks (CNNs), including Parallel network (PCNN), Soft Attention Convolutional Neural Networks (SPCNN), after Functional Block (SFPCNN), improve effectiveness of malaria diagnosis. Among these, SPCNN emerged most successful model, outperforming all other models evaluation metrics. The achieved precision 99.38 $$\pm$$ 0.21%, recall 99.37 F1 score accuracy ± 0.30%, an area under receiver operating characteristic curve (AUC) 99.95 0.01%, demonstrating its robustness detecting parasites. Furthermore, various transfer learning (TL) algorithms, VGG16, ResNet152, MobileNetV3Small, EfficientNetB6, EfficientNetB7, DenseNet201, Vision Transformer (ViT), Data-efficient Image (DeiT), ImageIntern, Swin (versions v1 v2). proposed model surpassed TL methods every measure. 2.207 million parameters size 26 MB, more complex than PCNN but simpler SFPCNN. Despite this, exhibited fastest testing times (0.00252 s), making it computationally efficient both We assessed interpretability using feature activation maps, Gradient-weighted Class Activation Mapping (Grad-CAM) SHapley Additive exPlanations (SHAP) visualizations for three architectures, illustrating why outperformed others. findings from our experiments show significant improvement parasite approach outperforms traditional manual microscopy terms speed. This study highlights importance utilizing cutting-edge technologies develop robust effective diagnostic tools prevention.

Language: Английский

---