Colon cancer stage detection in colonoscopy images using YOLOv3 MSF deep learning architecture

Colo-Rectal Cancer (CRC) stood witnessed as the major cause of deaths worldwide, especially in men. But the early detection and removal of polyps can reduce the
death rate to a considerable extent. The various stages of colon cancer using the YOLOv3 Multi-Scale Framework (YOLOv3-MSF) have been studied here in this
context. Colonoscopy is an examination method to detect changes in the rectum by inserting a long flexible tube into the rectum. For screening colon cancer,
gastroenterologist often refers to the pathological images of colonoscopy for diagnosis. In recent decades, various machine learning algorithms are used for diagnosis
to get practical and reliable results. More often, the stages of colon cancer can be inspected by the Tumor Length (TL). In this study, the polyps are segmented and
allowed to measure the TL value by assessing polyp length (Plen) and polyp width (Pwid). The various stages of colon cancer have been ascertained based on the
maximum length between the Plen and Pwid. Our research uses the single-stage YOLOv3 object detection model to locate and classify the various stages based on the
TL value. It outputs the class probability and position coordinates of each image by enhancing accuracy and robustness. At the outset, during preprocessing to address
the specific problems like reduced contrast, blurred images and noisy data, the proposed model defines a data augmentation with a jitter value of 0.3 to avoid
overfitting. Next, for detecting and annotating various stages of colon cancer, YOLOv3-MSF deep learning architecture with multiscale detection layers used here. By
adopting transfer learning, we used YOLOv3 in our work. It chooses the anchor boxes using the K-Medoids algorithm during object detection in the training stage
based on the ResNet network structure of YOLOv3. Finally, additional layers such as fully connected layer are embedded into the existing architecture. we used the
CVC colonDB database to train and evaluate the fully connected layer during experiments. To analyze the performance of the proposed system, we compare our
model with the state of art object detection model. Experimental results show that the proposed model outperforms existing systems on peak to accuracy, precision
rate, recall rate, and f-measures.