Axis-Wise Ridge Regression Blending of Heterogeneous 2D-3D Registration Methods for Cochlear Implant Surgery

Ziyi Wang, Hannah Mason, Jack Noble

Keywords: 2D-3D registration, cochlear implant surgery, surgical navigation, ridge regression, blending, pose estimation

TL;DR: We propose an axis-wise ridge regression blending method that improves translation accuracy for 2D-3D cochlear implant registration by combining five complementary deep learning models.

Registration Requirement: Yes

Abstract: Accurate registration between intra-operative 2D surgical microscope images and pre-operative 3D CT images is critical for image-guided cochlear implant surgery, as it enables augmented reality guided surgery. Five complementary deep learning methods have been developed for this registration task, but their architectural biases lead to uneven performance across translation and rotation axes, so no single method performs best on all DoFs. We propose an axis-wise ridge regression blending strategy that divides the five methods into two groups: direct pose regression methods (M1, M2) for in-plane translation, and affine-based methods (M3–M5) for depth and rotation. Separate ridge regression branches are trained for different output groups, with each branch taking the full 6-DoF predictions from its assigned method subset to preserve cross-axis information. Under leave-one-out cross-validation on seven patient samples, the proposed method achieves mean absolute errors of 0.5821, 0.2991, and 48.2001 mm on TX, TY, and TZ, respectively, outperforming all five baseline methods. The closed-form solution requires no GPU and remains stable with as few as six training samples, making it well suited to data-limited surgical navigation.

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 106