Enhancing repeatability of follicle counting with deep learning reconstruction high-resolution MRI in PCOS patients

Abstract

Follicle count, a pivotal metric in the adjunct diagnosis of polycystic ovary syndrome (PCOS), is often underestimated when assessed via transvaginal ultrasonography compared to MRI. Nevertheless, the repeatability of follicle counting using traditional MR images is still compromised by motion artifacts or inadequate spatial resolution. In this prospective study involving 22 PCOS patients, we employed periodically rotated overlapping parallel lines with enhanced reconstruction (PROPELLER) and single-shot fast spin-echo (SSFSE) T2-weighted sequences to suppress motion artifacts in high-resolution ovarian MRI. Additionally, deep learning (DL) reconstruction was utilized to compensate noise in SSFSE imaging. We compared the performance of DL reconstruction SSFSE (SSFSE-DL) images with conventional reconstruction SSFSE (SSFSE-C) and PROPELLER images in follicle detection, employing qualitative indices (blurring artifacts, subjective noise, and conspicuity of follicles) and the repeatability of follicle number per ovary (FNPO) assessment. Despite similar subjective noise between SSFSE-DL and PROPELLER as assessed by one observer, SSFSE-DL images outperformed SSFSE-C and PROPELLER images across all three qualitative indices, resulting in enhanced repeatability in FNPO assessment. These results highlighted the potential of DL reconstruction high-resolution SSFSE imaging as a more dependable method for identifying polycystic ovary, thus facilitating more accurate diagnosis of PCOS in future clinical practices.

Keywords: Deep learning; High resolution; Magnetic resonance imaging; Ovary; Polycystic ovary syndrome.

Fig. 1

Current challenges in follicle counting using traditional fast spin-echo T2-weighted MR imaging. Panel (A) shows the full field of view image, which offers relatively high temporal resolution. However, its low spatial resolution makes follicle counting challenging. In contrast, panel (B) presents a small field of view image with higher spatial resolution, but it is prone to noise and motion artifacts due to its lower temporal resolution, which limits accurate follicle counting.

Fig. 2

Scoring criteria for qualitative analysis. Scores of 3, 2, and 1 for blurring artifacts are assigned to (A–C), respectively. Scores of 3, 2, and 1 for subjective noise are assigned to (D–F), respectively. Scores of 4, 3, 2, and 1 for conspicuity of follicles are assigned to (G–J), respectively. Higher scores indicate better image quality.

Fig. 3

Representative T2-weighted MR images of the ovaries in a 17-year-old adolescent girl with PCOS. The SSFSE-DL images (A–G) clearly depict bilaterally enlarged ovaries and an increased number of peripheral follicles with minimal noise and blurring artifacts, providing the best conspicuity of follicles. The use of AIR Recon DL enhances the contrast between follicles and the surrounding ovarian stroma. The conspicuity of follicles in the SSFSE-C images (B, E, H) is mainly affected by noise, while that in the PROPELLER images (C, F, I) is primarily impaired by motion-related blurring artifacts.

Fig. 4

Representative T2-weighted MR images of the ovaries in a 24-year-old woman with PCOS. The SSFSE-DL images (A, D, G) clearly show bilateral ovaries and follicles with minimal noise and blurring artifacts, providing optimal conspicuity of follicles. In SSFSE-C images (B, E, H), follicle conspicuity is primarily influenced by noise, whereas in PROPELLER images (C, F, I), it is predominantly affected by motion-related blurring artifacts.

Fig. 5

Bland‒Altman plots for intra-observer and inter-observer differences of FNPO assessment. The 95% limits of agreement (LOA) for intra-observer (A) and inter-observer (B) variability on SSFSE-DL images are narrower than the those for intra-observer (C) and inter-observer (D) variability on SSFSE-C images, as well as the intra-observer (E) and inter-observer (F) variability on PROPELLER images.