Abstract:
Colon polyp is the growth of malignant polyps that are highly dangerous for women and men’s
lives. Colonoscopy based image analysis method is most commonly used to screen benign and
malignant polys. Recently the development of machine learning and deep learning, convolutional
neural networks create a great impact in the field of object detection due to the potential and
efficient extraction capabilities. Since it is a very strong and impactful technology because of this
it is widely used in medical image analysis and organ abnormalities for instance colorectal cancer.
Globally colorectal cancer is a great incident, but early detection of polyps decreases the mortality
rate. In this study, we employed a model that is responsible early detection of polyps. Firstly we
collect data from Kvasir-SEG data set and then apply the data augmentation technique to enhance
our dataset for better performance. In a consistent manner, 80% of data was conducted for training
and 20% of data for testing. From the training dataset, 20% of data was conducted for validation
for avoiding overfitting. After that, we used strong and very impactful to detect small objects
YOLOv5(You only look once) three versions YOLOv5s, YOLOv5m, YOLOv5l to investigate the
dataset and superintend their performance and compare each of them. In the training period, we
increase our iteration several times to enrich the model performance. After the final experiment,
YOLOv5l provides better performance than YOLOv5s and YOLOv5m. However, the YOLOv5l
acquire training accuracy, precision, recall, validation, and training loss and mean average
precision (mAP) are 97.00%, 85.00%, 0.0025, 0.035, and 85.1% gradually, which insure the better
detection performance of the model. Finally, the three versions of the YOLOv5 model are used to
evaluate the testing data with the various situations, conclude the colon polyps abnormalities
detection with the specific location are successfully detected.
