Enhancing Robustness in Multi-modal CT Image Segmentation with a Clustering-based Deep Active Learning Approach

Supervised deep learning techniques have become a powerful tool in many applications, including medical image segmentation. However, their efficacy depends heavily on the availability of large and uniformly annotated datasets. The need for high-quality annotations presents a significant challenge as it requires significant cost, time, and effort. This challenge is particularly relevant in medical imaging, where segmentation tasks require a high degree of accuracy and precision. To address this challenge, we propose a novel framework for active learning in multi-modal medical image segmentation. Our approach involves selecting optimal training samples from multi-modal Spectral CT medical images during the initial sample selection phases. Our experimental results demonstrate the effectiveness of our proposed framework, which exhibits higher precision and more robust outcomes than conventional active learning methods. Specifically, our proposed image training sample selection methods can achieve stable model performance for lesion segmentation using less than ten multi-modal medical image samples, compared to conventional single modal methods. Furthermore, as the number of annotated training samples increases, our approach continues to display superior accuracy and robustness compared to other initial sample selection methods. In summary, by leveraging the advantages of both representativeness and model uncertainty, our approach demonstrates improved precision and robustness, highlighting its potential as a valuable tool for lesion segmentation applications.