Pelvimetry is the study of the pelvis morphology in women for labor planning and medical assessment. This can be achieved by manually annotating pelvic CT images for extracting several measures of interest, which can be both time-consuming and subjective. While machine learning has achieved significant success in 2D landmarking applications, results in pelvic CT images are still limited, particularly with small datasets. This paper presents a two-step approach for detecting 3D landmarks in pelvic CT images. First, a simple CNN coarsely estimates landmark locations, serving as a starting point for further refinement. Then, higher resolution 3D patches and independent neural networks are used to obtain the final position for each landmark. Our model has shown promising results, obtaining an average distance error of 6.71 mm across 7 landmarks. These values allowed the calculation of the morphological measurements, demonstrating a strong correlation with the manual values. The proposed model has shown promising results, offering efficient and accurate predictions of the anatomical landmarks in CT examinations.
P. et al. (Mon,) studied this question.