Çoklu yaklaşimli histopatoloji görüntülerinin siniflandirilmasi

Histological images play a crucial role in diagnosing diseases, especially breast cancer, which remains a major health concern for women, and colon and lung for people worldwide. Computer-aided diagnosis tools significantly assist physicians in early detection and treatment planning, helping reduce mortality rates. Although, Convolutional neural networks (CNNs) based on deep learning have proven effective in distinguishing benign from malignant breast cancers, the scarcity of variability and feature diversity in small datasets would steer clear of CNNs and result in overfitting instances, as well as diminished discrimination. Deep learning is very sensitive to the hyperparameter optimization, and for the reason, the learning hierarchies could be the solution for network's comprehension of the relationship subtlety between the of the data types. In this context, this dissertation introduced a HAFMAB-Net: Hierarchical Adaptive Fusion based on Multilevel Attention-Enhanced Bottleneck Neural Network. The network comprises two pathways utilizing an enhanced Bottleneck architecture with attention mechanisms to extract both global and spatial features. It incorporates a Deeper Spatial Attention Aggregator Module to boost the representation of locative features by focusing on key spatial regions, improving the discriminative power of aggregated features. Additionally, a modified Adaptive Fusion Module combines the enhanced global and boosted spatial features into a comprehensive and enriched feature representation, which is subsequently used for cancer classification based on whole image. The histopathological image could include more than one type cancer inside one image, to address this issue, we proposed a novel MATERB Net: Multiscale Attention transformation based on Enhanced Residual Block for Patch-level breast histopathology images prediction. The proposed framework employs extraction feature processes based on multiscale to capture more complementary global and local characteristics to enhance the learning process of the model in coping with morphological variability within the tissue. In Additional the model uses Attention transformation techniques based on self-attention and cross-attention is used to learn the relationship between the pixels of the image and focus on the meaningful details and information in calculating the attention weights. Self-attention works on capturing and learning the relation between the extracted features based on multi-magnification factors sub-images, providing various levels of details and information, which feed into the cross-attention beside the extracted features from full images. Cross-attention works to provide meaningful and context-aware representation integrated with the global context and details of each label. Moreover, the proposed net used a Balanced Focal Loss function, addressing the treating all misclassified samples equally, and preventing the model from memorizing easy samples to reduce overfitting.