Effective Early Polyp Detection from Medical Images with YOLO-V7

Polyps constitute a potentially harmful cluster of cells appearing in the colon lining. Without early diagnosis and treatment, they can lead to Colorectal cancer (CRC), a quite common and highly lethal type of cancer. The early detection of a polyp, via a colonoscopy screening, can prevent CRC. However, the traditional colonoscopies exhibit high polyp-miss detection rates, leading to late CRC diagnoses. Therefore, studies for computer-aided detection systems for polyps are highly important. In our study, we employed the YOLOv7 model for the detection of polyps using transfer learning, utilizing a well-established endoscopic dataset. We take advantage of non-polyp images while training, to render our model more robust, achieve generalization, mitigate class imbalance and reduce noise. By training the model to recognize not only objects but also background elements, we reduce the risk of false positives. Compared to other researches, our model demonstrates a notable performance in both accuracy and real-time detection speed, achieving Precision and Recall values equal to 0.877 and mean average precision equal to 0.929.