3620
Automatic deep learning method for detection and classification of breast lesions in dynamic contrast-enhanced magnetic resonance imaging1The Second Affiliated Hospital of Xi’an Jiaotong University, Xi 'an, China, 2Northwest University, Xi 'an, China, 3GE HealthCare MR Research, Beijing, China
Synopsis
Keywords: Diagnosis/Prediction, Breast
Motivation: As manual slice-by-slice analysis of breast MR images is both time-consuming and error-prone.
Goal(s): To develop a deep learning-based system for the detection and classification of breast lesions in DCE-MRI.
Approach: DCE-MRI images were fed into the developed cascade feature pyramid network system(CFPN), feature pyramid network, and faster region-based convolutional neural network for breast lesion detection and classification.
Results: CFPN achieved the highest sensitivities in detection at the lowest FPs at both the slice level and the patient level.
Impact: DL-based systems can automatically detect and classify breast lesions on DCE-MRI. These results illustrate the potential use of this technique in a clinically relevant setting.
Introduction
Breast cancer is the most common cancer in women and the leading cause of cancer deaths in women worldwide1. Accurate, early diagnosis has an important impact on treatment planning and improving survival rates for breast cancer patients2. Dynamic contrast-enhanced (DCE)-MRI is the most sensitive imaging technique for the diagnosis of breast cancer3,4. Convolutional neural network (CNN) models have been widely used as a common Deep learning (DL) approach for breast MRI, including segmentation5, detection6, and classification of lesions7-9. However, considering its independent clinical application, DL needs to simultaneously detect and classify lesions in MRI, which has not been well investigated. The Faster region-based CNN (R-CNN)10 is a wellknown DL model that is based on a CNN with additional components for detecting, localizing, and classifying objects in images. Based on the Faster R-CNN, feature pyramid networks (FPNs) with multi-scale and pyramidal hierarchy architectures11 have also been used to detect breast masses by reducing the false positive rate of DL-based systems without significantly decreasing the detection sensitivity12,13. Furthermore, the DL-based system with cascade FPN (CFPN) has shown promising results in the detection of pulmonary nodules14. Therefore, the main goal of this study was to evaluate the performance of the CFPN system in detecting and classifying breast lesions in DCE-MRI compared to Faster R-CNN and FPN systems.Material and Methods
From January 2016 to July 2020, 191 patients ( 157 with malignant lesions and 34 with benign lesions) were retrospectively included in our study. All breast lesions were confirmed by biopsy or surgical pathology. DCE-MRI examination was performed on a 3.0 T system (Signa HDxt; GE Medical Systems, Milwaukee, WI, USA) with a dedicated 8-channel breast coil. Pre- and post-contrast phases were acquired before and after the injection of 0.2 mmol/kg body weight of gadolinium diethylenetriaminepentaacetic acid (DTPA) for 64 s, 128 s, 192 s, 256 s, and 318 s at a rate of 2.0 mL/s with a power injector followed by 20 mL saline solution, where the dose of gadolinium-DTPA followed. We used T1-weighted 3-dimensional (3D) fast spoiled gradient-recalled echo sequence with parallel imaging (VIBRANT) sequence on the transverse plane with the following parameters. The contrast-enhanced MRI acquired at 128 s was used for the detection of breast lesions as it had the best signal intensity contrast. Figure 1-2 shows the framework for DL-based analysis. The proposed DL-based workflow included the annotation of breast lesions, the augmentation of training data, and the detection and classification by different systems, including the Faster R-CNN, FPN, and CFPN systems. Student’s t-test was adopted to compare the performance of different systems in terms of the aforementioned metrics, followed by the Holm-Bonferroni method for adjusted P values in the comparison among multiple groups. AP value <0.05 was considered statistically significant.Results
The CFPN and FPN systems with the optimal settings (including augmentation and focal loss), as well as the classical Faster R-CNN were used to detect and classify benign and malignant breast lesions. As shown in Table 1, the AP values of CFPN for benign and malignant lesions were 0.731±0.075 and 0.921±0.048, respectively, which were both better than or comparable to the other 2 systems. The PR curves and FROC curves of median results in 5-fold cross-validation are shown in Figure 3. In particular, the CFPN achieved the highest sensitivities in detection at the lowest FPs at both the slice and patient level (Table 2).Discussion
In this study, we compared the DL-based CFPN system with the FPN and Faster R-CNN systems in the detection and classification of benign and malignant breast lesions, and further evaluated their performance for detecting large and small breast lesions. The results demonstrated that the mAP values and sensitivity of the CFPN outperformed the FPN and Faster R-CNN systems. CFPN achieved the highest sensitivities in detection at the lowest FPs at both the slice level and the patient level. This is crucial for the early follow-up treatment of patients with malignancy and for reducing overtreatment of patients without malignant lesions. A previous study by Zhou et al.6 showed that weakly supervised 3D DL could be used for breast cancer classification and localization in DCE-MRI. In contrast to their study, our study achieved simultaneous detecting, localizing, and classifying of the same lesion.Conclusion
To conclude, DL-based systems can automatically detect and classify breast lesions on DCE-MRI, and that CFPN has the highest sensitivity at the lowest FPs. These results illustrate the potential use of this technique in a clinically relevant setting.Acknowledgements
No acknowledgement found.References
1. Sung H, Ferlay J, Siegel RL, Laversanne M, Soerjomataram I, Jemal A, Bray F. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin 2021;71:209-49.
2. Byers T, Wender RC, Jemal A, Baskies AM, Ward EE, Brawley OW. The American Cancer Society challenge goal to reduce US cancer mortality by 50% between 1990 and 2015: Results and reflections. CA Cancer J Clin 2016;66:359-69.
3. Mann RM, Cho N, Moy L. Breast MRI: State of the Art. Radiology 2019;292:520-36.
4. Leithner D, Wengert GJ, Helbich TH, Thakur S, OchoaAlbiztegui RE, Morris EA, Pinker K. Clinical role of breast MRI now and going forward. Clin Radiol 2018;73:700-14.
5.Zhang L, Mohamed AA, Chai R, Guo Y, Zheng B, Wu S. Automated deep learning method for whole-breast segmentation in diffusion-weighted breast MRI. J Magn Reson Imaging 2020;51:635-43.
6.Zhou J, Luo LY, Dou Q, Chen H, Chen C, Li GJ, Jiang ZF, Heng PA. Weakly supervised 3D deep learning for breast cancer classification and localization of the lesions in MR images. J Magn Reson Imaging 2019;50:1144-51.
7.Truhn D, Schrading S, Haarburger C, Schneider H, Merhof D, Kuhl C. Radiomic versus Convolutional Neural Networks Analysis for Classification of Contrastenhancing Lesions at Multiparametric Breast MRI. Radiology 2019;290:290-7.
8.Dalmiş MU, Gubern-Mérida A, Vreemann S, Bult P, Karssemeijer N, Mann R, Teuwen J. Artificial IntelligenceBased Classification of Breast Lesions Imaged With a Multiparametric Breast MRI Protocol With Ultrafast DCE-MRI, T2, and DWI. Invest Radiol 2019;54:325-32.
9. Wan KW, Wong CH, Ip HF, Fan D, Yuen PL, Fong HY, Ying M. Evaluation of the performance of traditional machine learning algorithms, convolutional neural network and AutoML Vision in ultrasound breast lesions classification: a comparative study. Quant Imaging Med.Surg 2021;11:1381-93.
10. Ren S, He K, Girshick R, Sun J. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. IEEE Trans Pattern Anal Mach Intell 2017;39:1137-49.
11. Lin TY, Dollar P, Girshick R, He K, Hariharan B, Belongie S. Feature pyramid networks for object detection. 30th IEEE/CVF conference on computer vision and pattern recognition (CVPR); 2017, Jul 21-26; Honolulu, HI. IEEE, 2017.
12. Sajda P, Spence C, Pearson J. Learning contextual relationships in mammograms using a hierarchical pyramid neural network. IEEE Trans Med Imaging 2002;21:239-50.
13. Peng J, Bao C, Hu C, Wang X, Jian W, Liu W. Automated mammographic mass detection using deformable convolution and multiscale features. Med Biol Eng Comput 2020;58:1405-17.
14. Mu G, Chen Y, Wu D, Zhan Y, Zhou X, Gao Y. Relu Cascade of feature pyramid networks for CT pulmonary nodule detection. 10th International Workshop on Machine Learning in Medical Imaging (MLMI)/22nd International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI); 2019 Oct 13-17; Shenzhen, China. Springer, Cham, 2019.
Figures
