Neurofibroma detection using joint imaging genomics and ensemble learning

Abstract

Listen

Translate article

Accurate detection of neurofibromas is crucial for quantitative monitoring of tumor progression and surgical assessment. This paper proposes a method for neurofibroma detection in whole-body magnetic resonance imaging (WBMRI) using joint imaging genomics and ensemble learning. Firstly, we enhance texture features through a combination of image sharpening, filtering, brightness adjustment, and contrast enhancement. Then, we employ a weighted boxes fusion (WBF) technique based on test-time augmentation (TTA) under a single model and further integrate multiple models using the dual fusion approach of TTA and WBF. For segmentation, we utilize minimum bounding boxes based on segmentation masks for position calibration. Finally, false positive tumor regions are further eliminated through imaging genomics features. The experimental MRI data is obtained from collaboration between Harvard Medical School and domestic tertiary hospitals, comprising 158 cases with a total of 1380 tumors. Five-fold cross-validation is conducted with segmentation annotations completed by domain experts. Compared to the best results of single models, our proposed method achieves a 10.1% increase in average precision (AP), 7.8% increase in sensitivity, reduction of average false positives to 3.58, a decrease of 17.68, and an 8.5% improvement in competitive performance metric (CPM). This method effectively enhances the accuracy of neurofibroma detection and is applicable to detecting tumors and lesions in other medical imaging applications.