SaSaMIM: Synthetic Anatomical Semantics-Aware Masked Image Modeling for Colon Tumor Segmentation in Non-contrast Abdominal Computed Tomography

Pengyu Dai, Yafei Ou, Yuqiao Yang, Dichao Liu, Masahiro Hashimoto, Masahiro Jinzaki, Mototaka Miyake, Kenji Suzuki

Abstract: Colorectal cancer is a critical global concern, despite advancements in computer-aided techniques, the development of early-stage computer-aided segmentation holds substantial clinical potential and warrants further exploration. This can be attributed to the challenge for localizing tumor-related information within the colonic region of the abdomen when doing segmentation and that cancerous tissue remains indistinguishable from surrounding tissue even with contrast enhancement. In this work, a task-oriented Synthetic anatomical Semantics-aware Masked Image Modeling (SaSaMIM) method is proposed that leverages both existing and synthesized semantics for efficient utilization of unlabeled data. We first introduce a novel fine-grain synthetic mask modeling strategy that effectively integrates coarse organ semantics and synthetic tumor semantics in a label-free manner. Thus, tumor location perception in the pretraining phase is achieved by means of integrating both semantics. Next, a frequency-aware decoding branch is designed to achieve further supervision and representation of the Gaussian noise-based tumor semantics. Since the intensity of tumors in CT follows Gaussian distribution, representation in the frequency domain solves the difficulty in distinguishing cancerous tissues from surrounding healthy tissues due to their homogeneity. To demonstrate the proposed method’s performance, a non-contrast CT (NCCT) colon cancer dataset was assembled, aiming at early tumor diagnosis in a broader clinical setting. We validate our approach on a cross-validation of these 110 cases and outperform the current SOTA self-supervised method for 5% Dice score improvement on average. Comprehensive experiments have confirmed the efficacy of our proposed method. To our knowledge, this is the first study to apply task-oriented self-supervised learning methods on NCCT to achieve end-to-end early-stage colon tumor segmentation. Our codes are available at https://github.com/Da1daidaidai/SaSaMIM.