Deep Active Learning for Computer-Aided Detection of Nasopharyngeal Carcinoma in MRI Images
Xin Zhao, Guanghui Han, Haojiang Li, Zenghui Wei, Mengfan Yang · 2023
Early detection and treatment of nasopharyngeal carcinoma has an important impact on improving the survival rate of patients. Computer-aided detection based on deep learning methods can automatically detect the presence of nasopharyngeal carcinoma on patient magnetic resonance images (MRI), assisting in the assessment of tumor progression. However, large-scale annotation of MRI images is not feasible because it is time-consuming and burdens the healthcare system. This paper proposes a weakly supervised nasopharyngeal carcinoma detection method suitable for MRI images, which can obtain better detection performance with a small amount of labeled data. We first generate a pseudo-color version of the MRI image based on a multi-window sampling method, which preserves richer information and improves the information utilization of the image. Then, active learning and deep learning are combined to construct an active detection model for nasopharyngeal cancer, and the most representative image set is selected from a large-scale unlabeled set by using instance-level image uncertainty for further annotation by experts, which significantly reduces the demand for image annotation in deep network. The proposed method is verified on the MRI image set of 800 patients with nasopharyngeal carcinoma. The experimental results show that the resampling method based on multi-window settings can improve the performance of the classical depth detection model by 1.5 %, while the active detection model of nasopharyngeal carcinoma only uses 20 % labeled data to achieve 92.6 % of the performance of the deep learning detector trained with all samples. Good performance is obtained when the label set is small. Our active detection method for nasopharyngeal carcinoma can detect the lesion area of nasopharyngeal carcinoma with high accuracy without large-scale labeled data, which significantly reduces the sample labeling burden of doctors.