Instance Segmentation of Airway Anatomies Using Mask R-CNN Prompt Adaptation-SAM

Abstract

Accurate identification of key anatomy structures in airway intubation, the primary step in general anesthesia, is crucial for surgical success and patient safety. Achieving both object detection and segmentation in this context using deep learning technologies is challenging due to limited labeled data, especially for difficult intubation conditions with anatomical abnormalities, trauma, or tumors. This study proposes an efficient improvement of Segment Anything Model (SAM) through Mask R-CNN prompt and an adaption technology to achieve a competent performance on airway anatomy instance segmentation tasks. We first constructed a labelled dataset of 1000 samples from difficult intubation conditions. Compared to U-Net, Mask R-CNN, and DeepLab, our model improved the Intersection over Union (IoU) by 4.4%, 6.5%, and 6.6%, and Dice coefficient by 4.4%, 5.5%, and 6.1%, respectively. Using Parameter-Efficient Fine-Tuning (PEFT) with adapter modules, our model demonstrates significant enhancement in identification performance of airway anatomies, achieving a Dice coefficient of 97.3% and improving the IoU up to 90.3%. Notably, our model outperformed others when using fewer segmentation mask labels, with improvement more pronounced as the number of labels decreases. Fine-tuning on similar medical images from public datasets of different scenarios resulted an IoU of up to 86.0% and a Dice coefficient of up to 91.9%, comparable to results from fine-tuning on 200 airway samples. These findings demonstrate that our proposed Mask R-CNN prompt Adaptation-SAM approach can effectively enhance performance while reducing computational resources demands, making it well-suited for complex clinical applications such as intubation. This study also offers a promising framework for future medical instance segmentation tasks.