Distinguishing benign and malignant lesions on contrast-enhanced breast cone-beam CT with deep learning neural architecture search. Issue 142 (September 2021)

• Record Type: Journal Article
• Authors: Ma, Jingchen; He, Ni; Yoon, Jin H.; Ha, Richard; Li, Jiao; Ma, Weimei; Meng, Tiebao; Lu, Lin; Schwartz, Lawrence H.; Wu, Yaopan; Ye, Zhaoxiang; Wu, Peihong; Zhao, Binsheng; Xie, Chuanmiao
• Is Part Of: European journal of radiology. Issue 142(2021)
• Publication Date: 2021-09
• Subjects: Artificial Intelligence -- Breast -- Deep Learning -- Neural Networks -- Computed Tomography -- X-Ray; AI artificial intelligence -- AUC area under the ROC curve -- BCBCT breast cone-beam computed tomography -- CI confidence interval -- CNN convolutional neural network -- CE-BCBCT contrast-enhanced breast cone-beam computed tomography -- MG mammogram -- NAS neural architecture search -- NC-BCBCT non-contrast-enhanced breast cone-beam computed tomography -- ROC receiver operating characteristic -- US ultrasound; Medical radiology -- Periodicals; Radiology -- Periodicals; Radiologie médicale -- Périodiques; Medical radiology; Periodicals; 616.075705
• DOI: 10.1016/j.ejrad.2021.109878
• Languages: English
• ISSNs: 0720-048X

Abstract

Highlights: Neural architecture search automatically generated optimal convolutional neural network. Our model was tailored for breast cone-beam CT images. Preliminary results show that the new model was faster and more efficient than ResNet-50. The new model performed similarly to the radiologists' visual ratings. Abstract: Purpose: To utilize a neural architecture search (NAS) approach to develop a convolutional neural network (CNN) method for distinguishing benign and malignant lesions on breast cone-beam CT (BCBCT). Method: 165 patients with 114 malignant and 86 benign lesions were collected by two institutions from May 2012 to August 2014. The NAS method autonomously generated a CNN model using one institution's dataset for training (patients/lesions: 71/91) and validation (patients/lesions: 20/23). The model was externally tested on another institution's dataset (patients/lesions: 74/87), and its performance was compared with fine-tuned ResNet-50 models and two breast radiologists who independently read the lesions in the testing dataset without knowing lesion diagnosis. Results: The lesion diameters (mean ± SD) were 18.8 ± 12.9 mm, 22.7 ± 10.5 mm, and 20.0 ± 11.8 mm in the training, validation, and external testing set, respectively. Compared to the best ResNet-50 model, the NAS-generated CNN model performed three times faster and, in the external testing set, achieved a higher (though not statistically different) AUC, with sensitivity (95% CI) and specificity (95% CI) of 0.727, 80% (66–90%), and 60% (42–75%), respectively. Meanwhile, the performances of the NAS-generated CNN and the two radiologists' visual ratings were not statistically different. Conclusions: Our preliminary results demonstrated that a CNN autonomously generated by NAS performed comparably to pre-trained ResNet models and radiologists in predicting malignant breast lesions on contrast-enhanced BCBCT. In comparison to ResNet, which must be designed by an expert, the NAS approach may be used to automatically generate a deep learning architecture for medical image analysis.