Registration of Deformed Tissue with a Geometry-Contrastive Transformer Approach

Yeganeh Aliyari, Mahdi Tavakoli

Keywords: Geometry-Contrastive Transformer, Breast Cancer, Lumpectomy, Tissue Tracking

TL;DR: This study introduces the GC-Transformer, trained on FEM based simulations, to enable precise tracking of deformable breast tumours in the context of simulated lumpectomy

Abstract: Breast cancer affects over 2.3 million people worldwide each year. In its early stages, the disease is typically treated surgically through either mastectomy, involving complete removal of the breast, or lumpectomy, a breast-conserving procedure in which only the tumour and adjacent tissue are excised. Lumpectomy is often preferred because it preserves breast tissue and is associated with lower morbidity, including reduced pain, fewer infections, and less impact on quality of life. Despite these advantages, a major challenge in lumpectomy is achieving accurate tumour localization, as intraoperative tissue deformation can compromise surgical precision. Current localization methods, both wire and non-wire based, have proven clinically effective but remain invasive, frequently resulting in patient discomfort [1]. Computational approaches such as Finite Element Methods (FEMs) and neural networks have been investigated as non-invasive alternatives for tumour localization. While FEMs provide biomechanically accurate modelling of tissue deformation, their high computational cost limits real-time applicability. Neural networks, by contrast, offer faster performance but often struggle to generalize across patients with different breast geometries [2]. To address these challenges, we introduce the Geometry-Contrastive (GC) Transformer, a novel deep learning framework specifically designed to deliver accurate, generalizable, and real-time tracking of deformable tumours in dynamic breast tissue [3].

Submission Number: 5