Informatics in Medicine Unlocked (Jan 2023)

Real time mobile AI-assisted cervicography interpretation system

  • Siti Nurmaini,
  • Muhammad Naufal Rachmatullah,
  • Rizal Sanif,
  • Patiyus Agustiansyah,
  • Irawan Sastradinata,
  • Legiran Legiran,
  • Annisa Darmawahyuni,
  • Ade Iriani Sapitri,
  • Anggun Islami,
  • Firdaus Firdaus,
  • Bambang Tutuko,
  • Nur Muhammad Erji Ridho Lubis

Journal volume & issue
Vol. 42
p. 101360

Abstract

Read online

Cervicography visual inspection after acetic acid application (VIA) has been recognized as an alternative early screening in resource-limited settings, such as Indonesia. However, the accuracy of VIA results primarily relies on the examiner's expertise, and due to inadequate and comprehensive training of healthcare workers, the accuracy of VIA is diminishing. Our primary goal was to develop a real time mobile AI-assisted cervicography interpretation system empowered by lightweight model to promptly and autonomously determine precise VIA results. Our custom dataset comprises a substantial collection of 702 subjects from Dr Mohammad Hoesin General Hospital, Indonesia which were classified into two conditions: 418 with abnormal cervixes and 302 as a control. We conducted two experiments: one focused on the detection of the region of interest (RoI) of cervix, and the other on the segmentation of precancerous lesions. In this study, we utilize the object detection approach using the combined You Only Look Once (YOLO) framework. As a result, the proposed model achieves an exceptional mean average precision (mAP) of 99% for RoI cervix detection, while the segmentation of lesions achieves a mAP of 73% and an average intersection over union score of 40%. Furthermore, the model showcases an inference time of 10.4 ms, reflecting its efficiency in processing images and generating results swiftly. We also assessed the model with two oncologist consultants, and the results indicated a satisfactory agreement with a Kappa value of 0.838. The high Kappa value signifies a substantial level of agreement between the model's predictions and the assessments made by the oncologist consultants. This further validates the effectiveness and accuracy of the model in lesion segmentation and highlights its potential utility in clinical settings.

Keywords