[Paper Review] From Preoperative CT to Postmastoidectomy Mesh Construction: Mastoidectomy Shape Prediction for Cochlear Implant Surgery

Yike Zhang, Eduardo Davalos+3 more|arXiv (Cornell University)|Jan 7, 2026

Hearing Loss and Rehabilitation0 citations

TL;DR

The paper presents a hybrid self-supervised and weakly-supervised framework to predict the mastoidectomy shape from preoperative CT scans, achieving a mean Dice score of 0.72 and advancing 3D postmastoidectomy surface reconstruction for CI surgery.

ABSTRACT

Cochlear Implant (CI) surgery treats severe hearing loss by inserting an electrode array into the cochlea to stimulate the auditory nerve. An important step in this procedure is mastoidectomy, which removes part of the mastoid region of the temporal bone to provide surgical access. Accurate mastoidectomy shape prediction from preoperative imaging improves pre-surgical planning, reduces risks, and enhances surgical outcomes. Despite its importance, there are limited deep-learning-based studies regarding this topic due to the challenges of acquiring ground-truth labels. We address this gap by investigating self-supervised and weakly-supervised learning models to predict the mastoidectomy region without human annotations. We propose a hybrid self-supervised and weakly-supervised learning framework to predict the mastoidectomy region directly from preoperative CT scans, where the mastoid remains intact. Our hybrid method achieves a mean Dice score of 0.72 when predicting the complex and boundary-less mastoidectomy shape, surpassing state-of-the-art approaches and demonstrating strong performance. The method provides groundwork for constructing 3D postmastoidectomy surfaces directly from the corresponding preoperative CT scans. To our knowledge, this is the first work that integrating self-supervised and weakly-supervised learning for mastoidectomy shape prediction, offering a robust and efficient solution for CI surgical planning while leveraging 3D T-distribution loss in weakly-supervised medical imaging.

Motivation & Objective

• Motivate accurate mastoidectomy shape prediction to aid presurgical planning, robotic assistance, and intraoperative navigation in cochlear implant (CI) surgery.
• Develop a framework that eliminates manual labeling by leveraging self-supervised learning with postoperative CT data.
• Enhance robustness to noisy labels through a novel 3D T-Distribution loss in a weakly-supervised setting.

Proposed method

• Two-stage framework combining self-supervised learning from preoperative and postoperative CT pairs to predict mastoidectomy regions.
• A SegMamba-based network with a pretrained SAM-Med3D encoder to produce an inverted probability map of the mastoidectomy region on preoperative CT.
• A weakly-supervised refinement using weak labels derived from the self-supervised outputs and a 3D T-Distribution loss to robustly model irregular mastoidectomy geometry.
• Introduction of a 3D T-Distribution loss (based on the Student-t distribution) to improve robustness to noisy labels in 3D medical image segmentation.
• Evaluation against standard 3D segmentation models showing superior Dice and lower HD95 on test data.

Experimental results

Research questions

• RQ1 Can mastoidectomy shape be accurately predicted directly from preoperative CT without manual annotations?
• RQ2 Does a hybrid self-supervised and weakly-supervised approach improve robustness to noisy labels in 3D mastoidectomy segmentation?
• RQ3 Does the proposed 3D T-Distribution loss outperform traditional losses for weakly-supervised 3D medical imaging tasks?
• RQ4 How does the predicted postmastoidectomy surface enable downstream navigation and visualization for CI surgery?

Key findings

• Mean Dice score of 0.72 for predicting the complex mastoidectomy shape from preoperative CT.
• The proposed method achieves state-of-the-art performance on multiple metrics compared to UNet-, transformer-, and SwinUNETR-based models.
• 3D T-Distribution loss improves robustness to noisy/weak labels, outperforming CE, BCE, Focal, MSE, and MAE losses in ablation tests.
• Ablation shows the Mamba-based SegMamba architecture outperforms U-Net and transformer baselines in self-supervised learning.
• Table 2 shows the proposed method attains the highest Dice (0.721) and lowest HD95 (16.159) among competing methods.
• The framework enables reconstruction of postmastoidectomy surfaces from preoperative CT for improved intraoperative navigation.