Bridging AI and Healthcare: Deep Learning for Medical Image Analysis

Abstract

Colorectal cancer (CRC) remains one of the deadliest malignancies worldwide, necessitating advanced diagnostic tools for accurate tissue analysis and lesion localization. This research addresses two complementary aspects of CRC diagnosis through deep learning, beginning with histopathological colon image classification followed by polyp segmentation in colonoscopy images. In this thesis, we propose two novel deep learning models based on attention mechanisms for these tasks, which demonstrate exceptional performance, outperforming previous published works. For histopathological colon image classification, we first fine-tune three pretrained models such as Vgg19, DenseNet121, and MobileNetV2, then establish a classic CNN model, followed by proposing a novel deep convolutional neural network architecture. Our proposed model achieves exceptional performance of 97.66% accuracy in identifying malignant tissues among benign cases. For polyp segmentation in colonoscopy images, challenged by morphological variability, we first implemented a classical U-Net architecture before introducing our proposed U-Net that innovatively combines convolutional block attention module (CBAM) with squeeze and excitation (SE) blocks. Additionally, we fine-tuned the Meta’s segment anything model (SAM). Our proposed U-Net architecture employs dual attention mechanisms to simultaneously focus on what features are important through channel attention and where to focus through spatial attention, while SE blocks boost the most useful features across channels. The proposed U-Net achieves competitive performance with 91.93% Dice score and 85.07% IoU demonstrating effective feature learning despite the challenging colonoscopy environment. Together, these methodologies highlight the transformative role of AI in colorectal cancer diagnostics, from endoscopic polyp detection to microscopic tissue analysis, enabling earlier, more accurate CRC detection and leading to better patient outcomes.