Enhanced GI Tract Cancer Diagnosis Using CNNs and Machine Learning Models

Gastrointestinal cancer exhibits the greatest mortality rate among all cancers, at 35.4%.Endoscopy is one of the few methods for obtaining visuals of gastrointestinal tract lesions.Manual cancer detection is arduous.Deep learning can autonomously diagnose gastrointestinal tract lesions.Automation produces erroneous detection results.This study used the challenging Hyper-Kvasir dataset for training and validation purposes.The dataset undergoes first preprocessing with Brightness Preserving Histogram Equalization.Furthermore, processed datasets comprise training and validation sets.For segmentation, pretrained backbone-based U-Net architecture is used.The U-Net backbones include EfficientNet-B0, Efficient-Net-B7, and DenseNet201.The pre-trained models utilize Ima-geNet, so Hyper-Kvasir is employed for the fine-tuning of gastrointestinal tract segmentation.The optimal Intersection over Union (IoU) is 85.2% for the EfficientNet-B7 backbone inside the U-Net design.A custom convolutional neural network is employed to classify the hyper kvasir dataset.The suggested network derives profound features for classification using artificial neural networks.The proposed methodology surpassed state-of-the-art (SOTA) methods.