Cervix-Seg-Net: A Context-aware Attention mechanism for Cervical Lesions Segmentation using Colposcopic images

Smith K. Khare, Berit Bargum Booth, Victoria Blanes-Vidal, Lone Kjeld Petersen, Esmaeil S. Nadimi

Keywords: Digital colposcopy, deep learning, semantic segmentation, cervical lesion.

TL;DR: This paper explores an effective and accurate segmentation of high-risk cervical lesions using a novel Cervix-Seg-Net model.

Abstract: Colposcopic examination is the final stage of cervical cancer (CC) screening. Currently, a gynecologist expert performs colposcopy and manually marks high-risk cervical lesions. This process can be time-consuming and prone to inter- and intra-observer variations. As a result, there is an urgent need for an artificial intelligence-based solution for automated mapping of high-risk cervical lesions for effective CC screening. However, the current techniques are limited to classifying cervical lesions and identifying regions of interest. Therefore, this paper explores an effective and accurate segmentation of high-risk cervical lesions using a novel Cervix-Seg-Net model. The proposed model comprised a hybrid combination of an EfficientNet-based encoder to extract multi-scale, high-resolution contextual representations, pyramid scene parsing to capture global and regional context, attention mechanisms for adaptive weighting, and refined skip connections for noise reduction. This novel architecture addresses the challenges of limited contextual awareness and boundary ambiguity to provide precise delineation of high-risk cervical lesions. The Cervix-Seg-Net evaluation yielded dice scores of 0.847, 0.811, and 0.748 for training, validation, and test sets, respectively, using holdout and cross-validation techniques on a unique dataset with exact mapping of the cervix. The results indicate the model’s robust performance and effective lesion localization, making it a potential candidate for real-world clinical integration and decision support.

Primary Subject Area: Segmentation

Secondary Subject Area: Detection and Diagnosis

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 75