YoTransViT: A transformer and CNN method for predicting and classifying skin diseases using segmentation techniques

Informatics in Medicine Unlocked, Год журнала: 2024, Номер 47, С. 101495 - 101495

Опубликована: Янв. 1, 2024

Skin disease cases are becoming more common, and diagnosing these diseases in a clinic is never an easy task. A deep learning (DL) based model was previously used to diagnose skin-related disorders; this may help address problems. But lately, the transformer has become well-liked effective for computer vision applications such as image identification skin condition detection. Consequently, transformers (ViT), which have shown exceptional performance conventional classification tasks, were present study. The goal of self-attention enhance significance key elements while muting distracting ones. More specifically, YoTransViT, enhanced network, suggested. suggested ViT structures with augmentation segmentation methods applied ISIC 2019 dataset. We also assessed effectiveness our method using alternative model, swin (SwinViT). In addition some most well-known neural network (CNN) designs, MobileNetV2, InceptionResNetV2 DenseNet201 hence creating baseline framework through comparison. Eight different types dermatological issues Various metrics evaluation computed confirm method. results demonstrated efficiency methodology, approaches reach accuracy 99.97% precision 100%. remarkable system predicting potential be extremely important field vision. Lastly, we established put into practice web-based architecture called YoTransViT real-time prediction diseases.

Язык: Английский

---

Skin Lesion Classification Through Test Time Augmentation and Explainable Artificial Intelligence

Journal of Imaging, Год журнала: 2025, Номер 11(1), С. 15 - 15

Опубликована: Янв. 9, 2025

Despite significant advancements in the automatic classification of skin lesions using artificial intelligence (AI) algorithms, skepticism among physicians persists. This reluctance is primarily due to lack transparency and explainability inherent these models, which hinders their widespread acceptance clinical settings. The primary objective this study develop a highly accurate AI-based algorithm for lesion that also provides visual explanations foster trust confidence novel diagnostic tools. By improving transparency, seeks contribute earlier more reliable diagnoses. Additionally, research investigates impact Test Time Augmentation (TTA) on performance six Convolutional Neural Network (CNN) architectures, include models from EfficientNet, ResNet (Residual Network), ResNeXt (an enhanced variant ResNet) families. To improve interpretability models' decision-making processes, techniques such as t-distributed Stochastic Neighbor Embedding (t-SNE) Gradient-weighted Class Activation Mapping (Grad-CAM) are employed. t-SNE utilized visualize high-dimensional latent features CNNs two-dimensional space, providing insights into how group different classes. Grad-CAM used generate heatmaps highlight regions input images influence model's predictions. Our findings reveal enhances balanced multi-class accuracy CNN by up 0.3%, achieving rate 97.58% International Skin Imaging Collaboration (ISIC 2019) dataset. comparable to, or marginally better than, complex approaches Vision Transformers (ViTs), demonstrating efficacy our methodology.

Язык: Английский

---

Optimizing Convolutional Neural Network Impact of Hyperparameter Tuning and Transfer Learning

Advances in computational intelligence and robotics book series, Год журнала: 2025, Номер unknown, С. 301 - 326

Опубликована: Янв. 17, 2025

This chapter examines skin cancer, particularly melanoma, which has a high mortality rate, making early diagnosis essential. It explores how convolutional neural networks (CNNs) can improve melanoma detection, providing detailed technical analysis of hyperparameters and their impact on model performance. Strategies for tuning hyperparameters, including random search Bayesian optimization, are demonstrated. Using the HAM10000 dataset, assesses different hyperparameter settings accuracy, sensitivity, specificity. Issues like class imbalance addressed with data augmentation resampling. The optimization methods DenseNet121 MobileNetV2 accuracies to 85.65% 84.08%, respectively.

Язык: Английский

---

Deep pixel-wise supervision for skin lesion classification

Computers in Biology and Medicine, Год журнала: 2025, Номер 193, С. 110352 - 110352

Опубликована: Май 20, 2025

Язык: Английский

---

RvXmBlendNet: A Multi-architecture Hybrid Model for Improved Skin Cancer Detection

Human-Centric Intelligent Systems, Год журнала: 2024, Номер unknown

Опубликована: Сен. 9, 2024

Abstract Skin cancer, one of the most dangerous cancers, poses a significant global threat. While early detection can substantially improve survival rates, traditional dermatologists often face challenges in accurate diagnosis, leading to delays treatment and avoidable fatalities. Deep learning models like CNN transfer have enhanced diagnosis from dermoscopic images, providing precise timely detection. However, despite progress made with hybrid models, many existing approaches still challenges, such as limited generalization across diverse datasets, vulnerability overfitting, difficulty capturing complex patterns. As result, there is growing need for more robust effective that integrate multiple architectures advanced mechanisms address these challenges. Therefore, this study aims introduce novel multi-architecture deep model called "RvXmBlendNet," which combines strengths four individual models: ResNet50 (R), VGG19 (v), Xception (X), MobileNet (m), followed by "BlendNet" signify their fusion into unified architecture. The integration achieved through synergistic combination architectures, incorporating self-attention using attention layers adaptive content blocks. This used HAM10000 dataset refine image preprocessing enhance accuracy. Techniques OpenCV-based hair removal, min–max scaling, histogram equalization were employed quality feature extraction. A comparative between proposed "RvXmBlendNet" (CNN, ResNet50, VGG19, Xception, MobileNet) demonstrated highest accuracy 98.26%, surpassing other models. These results suggest system facilitate earlier interventions, patient outcomes, potentially lower healthcare costs reducing invasive diagnostic procedures.

Язык: Английский

---

Oral Jawbone Cystic Lesions' Intelligent Classification: Exploration and Practice of the Vision Transformer Model

2022 7th International Conference on Signal and Image Processing (ICSIP), Год журнала: 2024, Номер unknown, С. 712 - 716

Опубликована: Июль 12, 2024

Язык: Английский

---

A method for measuring hairline length and discriminating hairline recession grades based on the BiSeNet model

Measurement Science and Technology, Год журнала: 2024, Номер 36(1), С. 015705 - 015705

Опубликована: Окт. 18, 2024

Abstract Hair plays an important role in a person’s appearance. According to survey by the World Health Organization, approximately 70% of adults have scalp and hair problems. Doctors currently make hairline recession diagnoses based on loss criteria, but this approach is subjective. This paper proposes novel method for objectively assessing grades. First, Bilateral Segmentation Network model utilized obtain facial segmentation image. Second, utilizes connected components improve results. Next, labeling key points used extract part features eyebrow region calculate related values. Finally, judgment length grade realized combining these with camera calibration. In paper, front-face images 50 volunteers were collected determination. The results expert doctors compared method. showed 1.3 cm difference average about 80% similarity judgments. conclusion, using machine vision methods measure height provides objective repeatable

Язык: Английский

---

Addressing Challenges in Skin Cancer Diagnosis: A Convolutional Swin Transformer Approach

Deleted Journal, Год журнала: 2024, Номер unknown

Опубликована: Окт. 22, 2024

Skin cancer is one of the top three hazardous types, and it caused by abnormal proliferation tumor cells. Diagnosing skin accurately early crucial for saving patients' lives. However, a challenging task due to various significant issues, including lesion variations in texture, shape, color, size; artifacts (hairs); uneven boundaries; poor contrast. To solve these this research proposes novel Convolutional Swin Transformer (CSwinformer) method segmenting classifying lesions accurately. The framework involves phases such as data preprocessing, segmentation, classification. In first phase, Gaussian filtering, Z-score normalization, augmentation processes are executed remove unnecessary noise, re-organize data, increase diversity. phase we design new model "Swinformer-Net" integrating U-Net frameworks, define region interest. At final classification, segmented outcome input into newly proposed module "Multi-Scale Dilated Neural Network meets (MD-CNNFormer)," where samples classified respective classes. We use four benchmark datasets—HAM10000, ISBI 2016, PH2, Cancer ISIC evaluation. results demonstrated designed framework's better efficiency against traditional approaches. provided classification accuracy 98.72%, pixel 98.06%, dice coefficient 97.67%, respectively. offered promising solution segmentation supporting clinicians diagnose cancer.

Язык: Английский

---

DLT-Embryo: A Dual-branch Local feature fusion enhanced Transformer for Embryo multi-stage classification

Biomedical Signal Processing and Control, Год журнала: 2024, Номер 102, С. 107266 - 107266

Опубликована: Дек. 6, 2024

Язык: Английский

---