HLDnet: Novel deep learning based Artificial Intelligence tool fuses acetic acid and Lugol’s iodine cervicograms for accurate pre-cancer screening

Abstract Cervical cancer is one of the major causes of women’s death and infertility around the world, especially in developing and underdeveloped countries. Early screening for high-grade squamous intraepithelial lesion or above (HSIL+), including early stages of cervical cancer, is the key to improve the survival rate. Recently, Artificial Intelligence (AI) tools have been developed via deep learning methods to automatically detect biopsy regions based on colposcopy images (i.e., cervicograms). However, most existing AI tools only learn from the single-channel acetic acid cervicograms, which is known to have the caveat of inflated false positive detections. Since both acetic acid and Lugol’s iodine cervicograms are considered by physicians in clinic, it is important for AI tools to fuse both types of cervicograms to achieve desired accuracy. Thus, we develop a novel deep learning based AI tool, referred to as HLDnet (HSIL+ Detection Network), to fuse the information of both types of cervicograms to ensure satisfactory accuracy of detecting HSIL+ regions. The HLDnet framework fuses the detection results of the dual-channel target detection algorithm by adopting late fusion. Specifically, HLDnet uses the Intersection over Union (IoU) decision algorithm to evaluate the proportion of common lesion region detected on both types of colposcopy images, which will reduce false positive rate and improve detection accuracy of HSIL+ regions. By applying HLDnet to real clinical cervicograms of 400 training and validation patients with HSIL+ and additional 200 test patients (100 HSIL+ and 100 LSIL-), we obtained accuracy 0.86 (sensitivity 0.82 and specificity 0.90) with the test data, outperforming the respective accuracy 0.61 or 0.53 by single-channel detection using only acetic acid cervicograms or Lugol’s iodine. Overall, our AI tool HLDnet is expected to improve the accuracy of automatic pre-cervix-cancer screening, so that it can help increase the availability of pre-cervix-cancer screening to worldwide people, especially in developing and underdeveloped countries.