| Abstrakt: |
In computer vision, particularly in label categorization, attributing features such as color, shape, and tissue size to each category presents a formidable challenge. Dense features related to each category have been validated in recent studies and developed as a multi-label classification problem. Still, notable difficulties remain in (1) classifying attributes more extensively over different object categories, (2) correlating category vulnerability, (3) capturing features in one way, and (4) predicting category labels of a slide with a dense feature map. We have proposed a pre-trained ResNet101-based novel global–local convolution technique to resolve these issues. The proposed model has used ResNet101 as a backbone with additional convolutional, regularization, and dense layers. This technique has two methods to extract the most contributed histopathological slide features. The global descriptor has helped the model to identify the WSI global feature WGF(color, shape, tissue size). In contrast, the local feature extractor has WLF, which fuses the region of interest toward the slides category. After that, we combined WGL features (WSI to patch3×4) as an extension of the informed model to learn dense features of multi-label breast cancer categories. After that, the GLNET model uses a fine-tuning mechanism with informed learned to different categories of faded dense layers. Generally, the global–local blocks make sense of the WSI global feature while gaining the object-of-interest characteristic. The proposed model has used the global–local feature composition for each category of breast cancer. Our proposed model has improved accuracy on two benchmarks and challenging BreakHis and ICIAR2018-BachChallenge datasets for multi-label cancer category prediction. The model results stated in different evaluation matrices verify that the proposed model gains 2% accuracy compared to the existing classifier. Finally, through a series of experiments, we have demonstrated that the proposed model significantly improves accuracy in training on histopathological slides characterized by their complex nature. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink