<![CDATA[Breast cancer histopathological grading is important for estimating tumor aggressiveness and supporting treatment planning, yet manual assessment can be influenced by observer variability. This study comparatively evaluates two convolutional neural network architectures, ResNet50 and VGG16, for automated multi-stage breast cancer histopathology grading based on the Nottingham histologic grading categories: Grade I, Grade II, and Grade III. The task was formulated as image-level classification using 1,742 hematoxylin and eosin-stained histopathological images obtained from 150 patients at an Anatomical Pathology Laboratory. All images were resized to 224 × 224 pixels, converted into normalized numerical RGB values, and augmented using horizontal and vertical flipping. The dataset was divided into training, validation, and testing subsets while considering patient-level separation to reduce data leakage. VGG16 was trained using stochastic gradient descent with a learning rate of 0.001, a batch size of 32 for training and 64 for testing, and 60 epochs. ResNet50 was trained using stochastic gradient descent and CrossEntropyLoss for 200 epochs, with raw logits in the final layer because CrossEntropyLoss internally applies softmax. Experimental results showed that ResNet50 achieved higher training accuracy (0.9735), higher validation accuracy (0.6792), lower training and validation loss (0.0809), and slightly higher recall (0.7095) than VGG16. Both models obtained the same test accuracy of 0.7590. These findings indicate that ResNet50 demonstrated stronger learning behavior and slightly better sensitivity, although external validation and class-wise analysis remain necessary before clinical application. The study also provides a controlled baseline for future CNN-based decision-support research in digital breast pathology settings and practice.]]>
Copyrights © 2026
