Cer-ConvN3Unet: an end-to-end multi-parametric MRI-based pipeline for automated detection and segmentation of cervical cancer.

Abstract

Background: We established and validated an innovative two-phase pipeline for automated detection and segmentation on multi-parametric cervical cancer magnetic resonance imaging (MRI) and investigated the clinical efficacy.

Methods: The retrospective multicenter study included 125 cervical cancer patients enrolled in two hospitals for 14,547 two-dimensional images. All the patients underwent pelvic MRI examinations consisting of diffusion-weighted imaging (DWI), T2-weighted imaging (T2WI), and contrast-enhanced T1-weighted imaging (CE-T1WI). The deep learning framework involved a multiparametric detection module utilizing ConvNeXt blocks and a subsequent segmentation module utilizing 3-channel DoubleU-Nets. The pipeline was trained and tested (80:20 ratio) on 3,077 DWI, 2,990 T2WI, and 8,480 CE-T1WI slices.

Results: In terms of reference standards from gynecologic radiologists, the first automated detection module achieved overall results of 93% accuracy (95% confidence interval 92-94%), 93% precision (92-94%), 93% recall (92-94%), 0.90 κ (0.89-0.91), and 0.93 F1-score (0.92-0.94). The second-stage segmentation exhibited Dice similarity coefficients and Jaccard values of 83% (81-85%) and 71% (69-74%) for DWI, 79% (75-82%), and 65% (61-69%) for T2WI, 74% (71-76%) and 59% (56-62%) for CE-T1WI.

Conclusion: Independent experiments demonstrated that the pipeline could get high recognition and segmentation accuracy without human intervention, thus effectively reducing the delineation burden for radiologists and gynecologists.

Relevance statement: The proposed pipeline is potentially an alternative tool in imaging reading and processing cervical cancer. Meanwhile, this can serve as the basis for subsequent work related to tumor lesions. The pipeline contributes to saving the working time of radiologists and gynecologists.

Key points: An AI-assisted multiparametric MRI-based pipeline can effectively support radiologists in cervical cancer evaluation. The proposed pipeline shows high recognition and segmentation performance without manual intervention. The proposed pipeline may become a promising auxiliary tool in gynecological imaging.