Advancing Colorectal Cancer Tissue Classification by Integrating Deep Learning Algorithms with a Novel Hybrid Filter Approach

Abstract

Assessing colorectal cancer (CRC) histological images is crucial due to their intricate nature. Timely CRC detection greatly influences treatment outcomes and patient sur vival rates. Traditionally, pathologists diagnose CRC by manually inspecting colon tis sue images obtained from biopsies. However, with the emergence of digital pathology, automated tissue type recognition within these images becomes crucial. Present method ologies typically integrate textual features with classifiers or adopt transfer learning to classify diverse tissue types. Nonetheless, the varied tissue characteristics within histo logical images present notable classification hurdles. In our study, we introduce a comprehensive dataset comprising 5,000 histological images of human CRC, covering eight unique tissue types. Our approach involves proposing an optimal classification method by fine-tuning convolutional neural network (CNN) parameters and selecting the most appropriate optimizer. We utilize CNNs alongside a variety of filtration techniques with the aim of accurately classifying these tissues. Additionally, we present a novel hybrid filtration method that combines the properties of different filters to enhance image quality and optimize image classification. Notably, our innovative hybrid filtration approach, when coupled with the EfficientNet B7 model, achieves an unprecedented accuracy of 98.7%. This breakthrough marks a significant advancement in CRC detection literature, highlighting the potential of neural network models, particularly EfficientNet-B7 combined with our newly developed filter, for precise and efficient colorectal cancer classification. This methodology effectively distinguishes between healthy and diseased large intestine tissues