The invention provides a gastric colon polyp detection method based on
deep learning, and the method comprises the steps: building a Swinin-T-Faster joint detection model, combining the global
perception capability of Transform and the multi-scale
feature extraction advantages of a
convolutional neural network, and employing a layered window attention mechanism to strengthen the focusing of a polyp region; a PAFPN feature
pyramid module is innovatively introduced, efficient fusion of 384-dimensional deep semantic features and 192-dimensional shallow detail features is realized through up-sampling and channel splicing, and the
diameter of a tiny polyp (
diameter lt; 5 mm); and designing a three-level post-
processing mechanism: firstly, carrying out frame fusion based on the condition that the overlapping degree is greater than 0.85 and the confidence coefficient weight, secondly, filtering a false positive frame according to a clinical morphological rule, and finally, optimizing a
confidence score through dynamic weighting of the confidence coefficient and the overlapping degree. According to the method, the average precision mean value is 98.02% on a Kvasir
data set, the
false positive rate is reduced by 18% compared with that of a traditional model, real-time reasoning of 30 frames per second is achieved on edge equipment by means of the TensorRT quantification technology, the three technical bottlenecks of
small target missing detection, background interference misjudgment and clinical deployment
delay in medical
image analysis are effectively solved, and the medical
image analysis efficiency is improved. And a high-precision and high-robustness solution is provided for intelligent diagnosis of the
digestive endoscopy.