Abstract
Oesophageal cancer is a serious threat to human health and life due to its high incidence levels and mortality rates. Early detection and diagnosis are crucial. However, existing oesophageal cancer detection models are plagued with missed detections and false-positives, especially with small and irregular lesions. To address these challenges, a novel approach called JS-DETR has been proposed, which combines Joint position channel attention and Shape adaptation improvement loss with DEtection TRansformer (DETR). In JS-DETR, several key improvements are made to enhance the accuracy of automated oesophageal cancer detection and localization. First, the DETR backbone is reinforced by incorporating joint position channel attention, enhancing the model’s ability to learn and utilize crucial features effectively. Second, shape adaptation improvement loss is employed to refine the model’s regression loss function, resulting in more accurate predictions of the precise locations of oesophageal cancer targets. Finally, transfer learning is utilized to fine-tune the enhanced model by transitioning it from the COCO dataset to the oesophageal cancer barium swallow imaging dataset. The JS-DETR model achieves a precision of 76.9%, a recall of 79.4%, an average precision of 87.0%, and an F1-score of 78.1% in the experimental results. Compared to other currently popular object detection models. The JS-DETR model enables more precise detection and localization of oesophageal cancer, offering clinicians a more accurate means of oesophageal cancer detection. Our code is available at https://github.com/learningmuch/JS-DETR.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call