Effect of resize and oversampling on breast cancerclassification

Breast cancer is a common disease in women and when diagnosed at an early stage, survival increases significantly. The diagnostic process is based on the examination of tissue sections taken by biopsy by coloring them with dyes under a light microscope. It is possible to transfer the examined cross-sectional areas to a digital environment thanks to slide scanners: The glass slide is scanned line by line by a high-resolution sensor; the scanning software creates images by mosaicing these lines and allows the pathologist to zoom and scroll on the screen. However, the fact that tumor foci are distributed on a micrometer scale and the wide area scanned makes manual examination both time-consuming and increases the risk of missing. Therefore, there is a need for computer-aided diagnostic systems that can automatically detect and mark suspicious species, supporting the pathologist. Today, these diagnostic systems are used and are constantly evolving. The success of these systems depends on the correct training of data-based methods, especially deep learning. However, it is generally seen that non-cancerous samples are much more than cancerous samples in histopathological image sets. This class imbalance leads deep models to memorize the majority class; learning minority tumor samples becomes difficult and the possibility of incorrect cancer cell data detection increases. Moreover, since these raw images can be in the order of gigapixels, they increase memory consumption, increase processing time and make it difficult to use in practice when used directly in model training. To reduce the imbalance, oversampling methods balance the data distribution by artificially multiplying cancerous samples; while resizing methods reduce the training time and hardware requirements to a reasonable level by reducing the images to smaller scales. However, oversampling can add false patterns, and resizing can soften textural details, leading to loss of diagnostic information. This thesis aims to systematically investigate how compatible these two techniques are with each other and to what extent they can improve model accuracy when configured correctly. The open access breast cancer (IDC) histopathology dataset was used in the study. Balanced subsets were created using 50 × 50 pixel images labeled by expert pathologists, six different oversampling strategies, and six different resizing techniques. In addition, only resizing subsets were added, making a total of 42 subsets. The generated subsets were evaluated using widely adopted convolutional neural network architectures and adapted versions of the Vision Transformer-based models. The findings show that synthetically enriching the cancer class significantly increases the ability of the models to learn the correct detections; this gain is further strengthened when combined with smooth resizing kernels that preserve detail well. Thus, by applying appropriate data balancing and scaling strategies together, both convolutional networks and transform-based approaches have the potential to provide reliable and practical diagnostic aid.