LSCS-Net: A lightweight skin cancer segmentation network with densely connected multi-rate atrous convolution

Computers in Biology and Medicine, Journal Year: 2024, Volume and Issue: 173, P. 108303 - 108303

Published: March 18, 2024

The rising occurrence and notable public health consequences of skin cancer, especially the most challenging form known as melanoma, have created an urgent demand for more advanced approaches to disease management. integration modern computer vision methods into clinical procedures offers potential enhancing detection cancer . UNet model has gained prominence a valuable tool this objective, continuously evolving tackle difficulties associated with inherent diversity dermatological images. These challenges stem from diverse medical origins are further complicated by variations in lighting, patient characteristics, hair density. In work, we present innovative end-to-end trainable network crafted segmentation This comprises encoder-decoder architecture, novel feature extraction block, densely connected multi-rate Atrous convolution block. We evaluated performance proposed lightweight (LSCS-Net) on three widely used benchmark datasets lesion segmentation: ISIC 2016, 2017, 2018. generalization capabilities LSCS-Net testified excellent breast thyroid nodule datasets. empirical findings confirm that LSCS-net attains state-of-the-art results, demonstrated significantly elevated Jaccard index.

Language: Английский

---

Segmentation of skin lesion using superpixel guided generative adversarial network with dual-stream patch-based discriminators

Biomedical Signal Processing and Control, Journal Year: 2024, Volume and Issue: 94, P. 106304 - 106304

Published: April 13, 2024

Language: Английский

---

Boundary-Aware Gradient Operator Network for Medical Image Segmentation

IEEE Journal of Biomedical and Health Informatics, Journal Year: 2024, Volume and Issue: 28(8), P. 4711 - 4723

Published: May 22, 2024

Medical image segmentation is a crucial task in computer-aided diagnosis. Although convolutional neural networks (CNNs) have made significant progress the field of medical segmentation, convolution kernels CNNs are optimized from random initialization without explicitly encoding gradient information, leading to lack specificity for certain features, such as blurred boundary features. Furthermore, frequently applied down-sampling operation also loses fine structural features shallow layers. Therefore, we propose boundary-aware operator network (BG-Net) which (GConv) and mechanism (BAM) modules developed simulate remote dependencies between channels. The GConv module transforms into that can extract features; it attempts more images boundaries textures, thereby fully utilizing limited input capture representing boundaries. In addition, BAM increase amount global contextual information while suppressing invalid by focusing on feature weight ratios Thus, perception ability BG-Net improved. Finally, use multi-modal fusion effectively fuse lightweight U-shaped branch multilevel feature, enabling low-level spatial details be captured shallower manner. We conduct extensive experiments eight datasets broadly cover evaluate effectiveness proposed BG-Net. experimental results demonstrate outperforms state-of-the-art methods, particularly those focused segmentation.

Language: Английский

---

Tackling the class imbalanced dermoscopic image classification using data augmentation and GAN

Multimedia Tools and Applications, Journal Year: 2023, Volume and Issue: 83(16), P. 49121 - 49147

Published: Oct. 27, 2023

Language: Английский

---

Automatic melanoma detection using discrete cosine transform features and metadata on dermoscopic images

Journal of King Saud University - Computer and Information Sciences, Journal Year: 2024, Volume and Issue: 36(2), P. 101944 - 101944

Published: Feb. 1, 2024

Nowadays, raw dermoscopic images in melanoma detection do not have acceptable performance. Machine learning helps detect accurately. There are extensive studies classic and deep learning-based approaches for the literature. Still, they accurate or require high data. This paper proposes a hybrid mechanism automated on based Discrete Cosine Transform features metadata. It is composed of three steps. First, extra information/artifacts deleted; remaining pixels standardized processing. Second, reliability improved by Radon transform, data removed using Top-hat filter, rate increased Wavelet Transform. Then, number reduced Locality Sensitive Discriminant Analysis. The third step divides into test ones to create image-based models Finally, best model selected metadata-based model. Simulation results show that decision tree provides most practical improving accuracy sensitivity. Besides, comparison demonstrate our improves F-Value superior other approaches.

Language: Английский

---

LCAMix: Local-and-contour aware grid mixing based data augmentation for medical image segmentation

Information Fusion, Journal Year: 2024, Volume and Issue: 110, P. 102484 - 102484

Published: May 18, 2024

Medical image segmentation often faces challenges related to overfitting, primarily due the limited and complex training samples. This challenge prompts use of self-supervised learning data augmentation. However, requires well-defined hand-crafted tasks multiple stages. On other hand, basic augmentation techniques like cropping, rotation, flipping, effective for natural scene images, have efficacy medical images their isotropic nature. While regional dropout regularization methods proven in recognition tasks, application is not as extensively studied. Additionally, existing operate on square regions, leading loss crucial contour information. particularly problematic dealing with regions interest characterized by intricate shapes. In this work, we introduce LCAMix, a novel approach designed segmentation. LCAMix operates blending two masks based superpixels, incorporating local-and-contour-aware strategy. The process augmented adopts auxiliary pretext tasks: firstly, classifying local superpixels using an adaptive focal margin, leveraging ground truth prior knowledge; secondly, reconstructing source mixed mutual masks, emphasizing spatial sensitivity. Our method stands out simple, one-stage, model-agnostic, plug-and-play solution applicable various tasks. Notably, it no external or additional models. Extensive experiments validate its superior performance across diverse datasets codes are available at https://github.com/DanielaPlusPlus/DataAug4Medical.

Language: Английский

---

Semi-supervised skin cancer diagnosis based on self-feedback threshold focal learning

Discover Oncology, Journal Year: 2024, Volume and Issue: 15(1)

Published: May 22, 2024

Worldwide, skin cancer prevalence necessitates accurate diagnosis to alleviate public health burdens. Although the application of artificial intelligence in image analysis and pattern recognition has improved accuracy efficiency early diagnosis, existing supervised learning methods are limited due their reliance on a large amount labeled data. To overcome limitations data labeling enhance performance diagnostic models, this study proposes semi-supervised model based Self-feedback Threshold Focal Learning (STFL), capable utilizing partial scale unlabeled medical images for training models unseen scenarios. The proposed dynamically adjusts selection threshold samples during training, effectively filtering reliable using focal mitigate impact class imbalance further training. is experimentally validated HAM10000 dataset, which includes various types lesions, with experiments conducted across different scales samples. With just 500 annotated samples, demonstrates robust (0.77 accuracy, 0.6408 Kappa, 0.77 recall, 0.7426 precision, 0.7462 F1-score), showcasing its Further, comprehensive testing validates model's significant advancements efficiency, underscoring value integrating This offers new perspective processing contributes scientific support treatment cancer.

Language: Английский

---

An artificial intelligence model for the semantic segmentation of neoplasms on images of the skin

Biomedical Engineering, Journal Year: 2024, Volume and Issue: 58(1), P. 36 - 39

Published: May 1, 2024

Language: Английский

---

Investigating the Quality of DermaMNIST and Fitzpatrick17k Dermatological Image Datasets

Scientific Data, Journal Year: 2025, Volume and Issue: 12(1)

Published: Feb. 1, 2025

Language: Английский

---

Advancing Non-Invasive Melanoma Diagnostics with Deep Learning and Multispectral Photoacoustic Imaging

Published: Jan. 1, 2025

Language: Английский

---