APPLICATION OF CONVOLUTIONAL NEURAL NETWORKS AND VISION TRANSFORMERS IN CANCER GRADING IN PATHOLOGY IMAGES

Abstract

Cancer Grading is a time-consuming and labor-intensive process. There is a need for accurate and robust Machine Learning (ML) models for automated Cancer Grading in Pathology Images. Existing methods use Convolutional Neural Networks (CNNs) for image classification and attention modules like Convolutional Block Attention Module (CBAM) for intermediate feature map refinement. However, integrating the Original Sequential CBAM between Convolutional Blocks in CNNs can disrupt the information flow in a model, increases the number of parameters, and can lead to longer and more computationally intensive training; our experiments demonstrate this can negatively impact performance. We propose Post-Convolutional Parallel CBAM for Cancer Grading in Pathology Images. We used KBSMC colon cancer dataset for training and validation for 20 epochs on three different architectures: VGG16, GoogLeNet, and ResNet34. The results indicate that the Proposed Post-Convolutional Parallel CBAM consistently outperforms Baseline and Original Sequential CBAM methods across various evaluation metrics despite resulting in fewer parameters than models using the Original CBAM integration. For example, the proposed method resulted in F-1 score of 0.810, while the Original CBAM approach got 0.658. Therefore, the proposed approach showed its effectiveness for transfer learning scenarios, and further development may lead to accurate and robust diagnostic tools.