Classification of breast lesions on DCE-MRI Data Using Fine-tuned MobileNet

doi:10.21203/rs.3.rs-1764475/v2

Download PDF

Research Article

Classification of breast lesions on DCE-MRI Data Using Fine-tuned MobileNet

https://doi.org/10.21203/rs.3.rs-1764475/v2

This work is licensed under a CC BY 4.0 License

Version 2

posted

You are reading this latest preprint version

Background

Early and rapid diagnosis of breast cancer is very important. Traditional method for detecting and diagnosing breast cancer may lead to a false positive or negative result. Presently, most of the deep learning models used in breast cancer detection prevents their use on mobile phones or low-configuration devices. This study intends to evaluate the capability of MobileNetV1 and MobileNetV2 and their fine-tuned models to differentiate malignant from benign lesions in breast dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI).

Methods

The diagnostic efficiencies of MobileNetV1_False(V1_False), MobileNetV1_True (V1_True), MobileNetV2_False(V2_False) and MobileNetV2_True(V2_True) were tested and evaluated under the same program architecture and dataset (2124 images in benign group and 2226 images in malignant group). A relatively optimal model was selected (with the highest accuracy in the training set and test set), and two fine-tuning strategies (S0 and S1) were used to modify it. The accuracy (Ac) of different models in the training and test sets, as well as the precision (Pr), recall rate (Rc), f1 score(f1) and area under the receiver operating characteristic curve (AUC), were taken as the performance indicators.

Results

The Ac of V1_True (0.9815) was higher than those of V1_False (0.9749), V2_False (0.9672) and V2_True (0.9699). Overfitting were observed in all models. Pr, Rc, f1 and AUC of V1_True were 0.79,0.73,0.74 and 0.74, respectively, they were higher than those of V1_False (0.77,0.73,0.73 and 0.73), V2_False (0.74,0.65,0.66 and 0.65) and V2_True (0.76,0.67,0.68 and 0.67).

Conclusion

The MobileNetV1_True model can differentiate between benign and malignant breast lesions on breast DCE-MRI, and has the potential to be deployed on an embedded mobile platform. Future work is necessary to improve the generalization capability of the proposed method.

Mobile convolutional neural networks

Deep learning

Breast lesions

Magnetic resonance imaging

In 2020, female breast cancer has surpassed lung cancer as the most commonly diagnosed cancer[1, 2]. It is also one of the leading causes of death for women[3, 4]. Breast cancer must be detected early and correctly diagnosed as malignant or benign in order to avoid its further progression and complications. By doing so, it is possible to plan an effective and timely treatment process, which in turn lowers the mortality rate associated with this disease.

A traditional method for detecting and diagnosing breast cancer have several important limitations[5, 6]. For example:

1. Experts are unavailable in remote areas (underdeveloped countries).

2. Lack of domain experts capable of accurately analyzing multiclass images.

3. The process of reviewing large numbers of medical images every day can be exhaustive and tedious.

4. As a result of the subtle nature of breast tumors and the complexity of breast tissue, manual examination is more challenging.

5. The concentration level of medical experts and other fatigues make diagnosing more difficult and time consuming.

The presence of such facts lengthens the diagnosis and may lead to a false positive or negative result. Adding additional methods to increase efficiency and reduce false predictions is always necessary. In recent years, artificial intelligence (AI) technology has made great progress in automatically analyzing medical images for anomaly detection. In comparison with manual inspection, automated image analysis using AI reduces the time and effort needed for manual image screening, and captures valuable and relevant information more efficiently from massive image collections[6–9].

Recent advances in AI research are focusing on the automation of early detection and diagnosis in light of the importance of early detection and diagnosis. In addition to this, the advent of robust AI algorithms (deep learning) has led to a surge in research activities in this field, it is important to have hardware that can train/run those robust and complex AI algorithms as well as large enough datasets for training AI algorithms. There are several imaging modalities that researchers have utilized to automate breast cancer detection, including mammography, ultrasound, magnetic resonance and histopathological imaging or any combination of them[3, 10–14].

Deep learning techniques achieved state of the art results in many computer vision tasks, and the medical field is no exception[15–17]. Presently, most of the deep learning models used in breast cancer detection are classic, the excessive number of parameters, which occupy considerable space, prevents their use on mobile phones or low-configuration devices, so their deployment on low-cost mobile computers is still challenging.

MobileNet is a lightweight model designed specifically for mobile and embedded terminals, compared with other convolutional neural networks, it reduced the number of parameters greatly[18]. In this article, we evaluated the capability of MobileNetV1 and MobileNetV2 and their fine-tuned models to differentiate malignant from benign lesions in breast DCE-MRI.

Data set

A total of 310 patients with complete DCE-BMRI and pathological data were collected from January 2017 to December 2020, including 17 patients with bilateral lesions (benign and malignant lesions on one side). All lesions were pathologically confirmed (surgical or needle biopsy) in this part. Lesions were divided into benign and malignant groups. Age, pathological type, and tumor diameter were compared between the two groups (Table 1). There was no significant difference in age between the two groups (P = 0.52). A statistically significant difference was observed in lesion diameter (p < 0.05). The inclusion criteria were as follows:Ⅰ: patients who did not receive any preoperative chemotherapy or chemoradiotherapy before MRI. Ⅱ: No puncture or surgical procedure was performed before the MRI examination. We discarded images under the condition that benign and malignant lesions appeared in the same image.

To avoid the influence of benign and malignant lesions in the bilateral breasts of the same patient, DCE-BMRI images were unilateral images one by one. Finally,2124 benign images (benign group) and 2226 malignant images (malignant group) were obtained. The images of the benign and malignant groups were randomly divided into a training set (benign group:1704 images, malignant group: 1786 images), test set (benign group:210images, malignant group:220 images) and validation set (benign group:210images, malignant group:220 images) according to a ratio of 8:1:1.

MRI Techniques

MRI was performed using two 3T MRI scanners with a dedicated breast coil in the prone position. Gd-DTPA (0.1 mmol/kg, 2.50 mL/s) was administered via elbow vein injection. A total of six phase images were acquired (one pre-contrast phase and five post-contrast dynamic enhancement phases). The detailed scanning parameters are listed in Table 2. MRI was performed preoperatively and before therapy initiation.

Proposed model

The computer environment was configured as the Windows 10 enterprise version of the 64-bit operating system, Intel (R) Core (TM) i7-10700F, NVIDIA RTX 2060 GPU, 6 GB. The other running programs are closed when the model is running. To facilitate the comparison of the network performance, the same training and validation set were selected for each network. The thresholds of the models were set at 0.5; if a result was ≥0.5, the image was judged as malignant, while it was predicted as benign otherwise. The architecture of the proposed DTL model for breast lesion classification is shown in Figure 1.

Data augmentation

The images were randomly shuffled using a set of programs. Data augmentation was performed before model training. The original images were augmented by flipping and rotating the original image, so that the augmented image maintained the original medical characteristics. The parameters and values are listed in Table 3.

Network structure of MobileMetV1

The main principle of a MobileNetV1 model is the application of depth-wise separable convolutions, which is made of a depth-wise convolution and a pointwise convolution[19]. The kernel of the depth-wise convolutional layer slides only to convolve with one input channel. Pointwise convolutional layer with a convolution kernel size of 1 × 1. Channels of output matrices are equal to the number of convolutional kernels, while channels of input matrices are equal to the number of convolutional kernels channels.

Network structure of MobileNetV2

In MobileNetV2, convolutional layers, bottleneck layers, and an average pooling layer form the basic network structure. The structure of the bottleneck layers can be found in reference[18], it usually includes pointwise convolution and depth-wise convolution; when the stride is 1, the input is added to the output. Another structural feature of MobileNetV2 inverted residuals[18]. In addition, a Relu6 serves as activation function in inverted residuals, this function is defined according to the expression:

$$\text{y}=\text{R}\text{e}\text{L}\text{U}\left(6\right)=\text{m}\text{i}\text{n}\left(\text{max}\left(\text{x},0\right)6\right)$$

Fine-tuning strategy

In this study, we designed two fine-tuning strategies for MobileNetV1 and MobileNetV2, they were S0 and S1. In S0, all parameters were non-trainable except for parameters in fine-tuned fully connected layers, while in S0, all trainable parameters were activated and along with parameters in fine-tuned fully connected layers were participated in the model training (Fig. 2). In this way, four models were generated for our study, they are MobileNetV1_False(V1_False), MobileNetV1_True(V1_True), MobileNetV2_False(V2_False) and MobileNetV2_True(V2_True).

The network structure is not changed throughout the training process. We selected the convergence of parameters and capacity for generalization as the primary outcome measure for the DTL models.

Hyperparameter settings

We used binary cross-entropy as our loss function. To explore suitable hyperparameter combinations for DTL workflows, we trained an DTL model for each classification task and each hyperparameter combination. (Fig. 3). The input image size was 224×224. Training for our model on DCE-BMRI images required 60 epochs with a batch size of 64 images. In addition, the activation functions were ReLU and sigmoid in fine-tuned fully connected layers, as shown in equations 1 and 2:

$$\text{R}\text{e}\text{L}\text{U}\left(\text{x}\right)=\text{f}\left(\text{x}\right)=\left\{\begin{array}{c}max(0,x), x\ge 0\\ 0, x<0\end{array}\right.$$

$$\text{s}\text{i}\text{g}\text{m}\text{o}\text{i}\text{d}\left(\text{x}\right)=\text{f}\left(\text{x}\right)=\frac{1}{1+{\text{e}}^{-\text{x}}}$$

Evaluation metrics

To compare the performance of DTL models, the following five performance indices were calculated: accuracy (Ac), precision (Pr), recall rate(Rc), F1 score ($\text{f}1$), and area under the receiver operating characteristic (ROC) curve (AUC) as the metrics in this study[20]. The positive and negative cases were assigned to the malignant and benign groups, respectively. Hence, true positive (TP) and true negative (TN) represent the number of correctly diagnosed malignant and benign lesions, respectively. FP and FN indicate the numbers of incorrectly diagnosed malignant and benign lesions, respectively. TP was considered a true-positive sample. This was also a positive sample. The mathematical formulations of Ac, Pr, Rc,$\text{f}1$are as follows:

$$\text{A}\text{c}=\frac{\text{T}\text{P}+\text{T}\text{N}}{\text{T}\text{P}+\text{T}\text{N}+\text{F}\text{P}+\text{F}\text{N}}$$

$$\text{P}\text{r}=\frac{\text{T}\text{P}}{\text{T}\text{P}+\text{F}\text{P}}$$

$$\text{R}\text{c}=\frac{\text{T}\text{P}}{\text{T}\text{P}+\text{F}\text{N}}$$

$$\text{f}1=\frac{2\times \text{P}\text{r}\times \text{R}\text{c}}{\text{P}\text{r}+\text{R}\text{c}}$$

The distribution of the data was not considered in Ac. $\text{f}1$ is a balanced metric determined by precision and recall; it is useful when there are imbalanced classes in the dataset.

The false positive rate (FPR) was calculated by dividing the total number of negatives by the fraction of negatives incorrectly classified as positive by the model. It can be evaluated as:

$$\text{F}\text{P}\text{R}=\frac{\text{F}\text{P}}{\text{F}\text{P}+\text{T}\text{N}}$$

A false negative rate (FNR) is the fraction of positives misclassified by the model as negative divided by the total number of positives. It can be evaluated as:

$$\text{F}\text{N}\text{R}=\frac{\text{F}\text{N}}{\text{F}\text{N}+\text{T}\text{P}}$$

Learning curves

We found that the combination of average pooling with all other hyperparameters (Adam, learning rate = 0.001 and dropout) perform best. By analyzing the metrics of models (V1_False, V1_True, V2_False, V2_True) on dataset, we found that the accuracy of all models reached 1.00 on the training set, and the highest accuracy(V1_True) was 0.9815 on test set, which was higher than V1_False (0.9749), V2_False (0.9672) and V2_True (0.9699), respectively.

Although the network architectures of MobileNetV1 and V2 were highly similar, the highest test accuracy of V1 was higher than that of V2(Fig. 4 and Fig. 5). However, they have the same disadvantages as the loss value has a trend of rising with the increase in epochs, which revealed that they did not converge on dataset and overfitting problem occurred. but the loss value was the lowest in the test set for V1_True, this reveal that the V1_True model have a better generalization ability than the other models. Visualizations of the training process (heat maps) are shown in Fig. 6.

The training time and model size

Total parameters of MobileNetV1 were more than that of MobileNetV2, while the time taken for training the model of MobileNetV2 were more than that of MobileNetV1, but the size of the file saved of MobileNetV2 were small than that of MobileNetV1. For detailed data, see Table 4.

Cross validation

Five-fold cross-validation was employed to examine the model performance. The five-fold cross-validation method and results are illustrated in Figure 7 and Table 5.

Classification report

The overall Pr, Rc, f1 and AUC of V1_True in validation set were 0.79,0.73,0.74 and 0.74, respectively. They were higher than those of V1_False, V2_False and V2_True. Detailed information is provided in Table 6 and Figure 8.

Visualization of confusion matrices

To intuitively show the superiority of the proposed DTL model, the confusion matrixes of the DTL models are given in Figure 9. FPR and FNR values of V1_True were 9.52% and 43.18%, respectively. FPR and FNR values of V1_False, V2_False, V2_True were 12.38%, 13.81%, 11.90% and 11.00%, and 41.36%, 55.91%,53.18, respectively. V1_True provided less FPR than other models. FNR of V1_True was slightly higher than that of V1_False, but lower than those of V2_False and V2_True.

Breast cancer is now the most commonly diagnosed cancer among women, with an estimated 2.3 million new cases each year, it is the fifth most common cause of cancer-related death worldwide. Among women, breast cancer accounts for 1 in 4 cancer cases and for 1 in 6 cancer deaths, ranking first for incidence in the vast majority of countries, and death rate for female breast cancer was considerably higher in transitioning versus transitioned countries[1]. Therefore, it is important to search for a method that can rapidly and efficiently diagnose early breast lesions.

In contrast to normal magnetic resonance imaging (MRI), Dynamic Contrast Enhanced MRI (DCE-MRI) provides a more detailed view of the soft breast tissues, so the affected breast areas can be easily identified. Literature reports concluded that diagnosis was the most commonly performed medical task, and deep learning techniques were most often used to perform classification, but the majority of selected studies used mammograms and ultrasound rather than magnetic resonance imaging as imaging modalities[3, 4, 21, 22].

Due to the memory and computation demands of deep neural networks, they are difficult to apply them to embedded systems with limited hardware resources, for instance, ResNet, VGG and DenseNet. Models like those listed above have many parameters and take up a lot of space. That is, the networks of these models almost always have deep structures, which are adequate for extracting complex image features and should be compressed and accelerated. By applying depth-wise separable convolutions, the MobileNetV1 is a lightweight model, which can decrease the number of parameters and computational complexity with less loss of classification precision[23]. In this way, the overall computation time could be reduced significantly, and has a relatively high diagnostic accuracy for the medical image classification task.

Arora etal.[24] analyzed two publicly available datasets (COVID-CT scan and SARS-COV-2 CT-Scan) retrospectively, they found that the MobileNet model provided a precision of 94.12% and 100% respectively. A separate study aimed at evaluating the effectiveness of the state-of-the-art pretrained convolutional neural networks on the automatic diagnosis of COVID-19 from chest X-rays (CXRs), the outcome of their study showed that MobileNet obtained the highest accuracy of 98.28%[25].

Encouraged by these results, we evaluated the capability of MobileNetV1 and MobileNetV2 and their fine-tuned models to differentiate malignant from benign lesions in breast DCE-MRI. We found that the MobileNetV1 models have a better generalization ability than MobileNetV2 models, and V1_True model obtained the highest accuracy on test set. But weaknesses of V1_True was that Pr and AUC only have 0.79 and 0.74, this illustrates that the robustness and generalizability of the proposed model need to be further increase in future studies. In other words, two important questions now before us are as following: whether the performance of proposed model can be further improved and how to improve it. An additional deficiency of this study was overfitting problem was observed in V1_True model. The loss value has a trend of rising with the increase in epochs, which revealed that it did not converge on dataset and overfitting problem occurred. We considered there are two main directions to solve this problem: expand the training data sets and optimize the model. More importantly, we observed that the proposed model only takes about 20 min to train the data, and the size of the model saved only have 19.4MB. This suggests that the proposed model was easy to deploy in the mobile terminal.

Several limitations to this study need to be acknowledged. Firstly, the number of images in training set was relatively small, especially lack of some rare lesions. Our dataset used for training may not represent the entire population of breast disease, this may have an impact on the accuracy of DTL model. Therefore, further analysis with more datasets is necessary to fully test the robustness of the DTL model. Secondly, during routine diagnostic procedures, the clinical evaluation, breast ultrasound and mammography are performed in addition to magnetic resonance imaging, however, only DCE-MRI images were adopted in our study. Thirdly, the high performance achieved by our proposed model was based on the premise of high-quality DCE-MRI images. In clinical practice, poor-quality images from other hospitals might decrease the performance of the DTL model. Therefore, High quality DCE-MRI images obtained by the standard procedure is highly warranted. A future study might require a multicenter collaboration to obtain a sufficiently large series of data to train and test proposed neural network.

Using deep learning techniques, we proposed an intelligent lightweight-assisted model to differentiate benign and malignant lesions, called MobileNet, based on our datasets with data augmentation. Results were compared between four models, and V1_True model achieved the best performance. Low-configuration computers are compatible with this model, this is particularly advantageous for some units where the computer conditions are limited. A suitable recommendation and timely referral can also be provided by it, obtain a high level of diagnosis for imaging diagnosis and result in a good social benefit. Additionally, in future, this lightweight model could be embedded in mobile devices so that mobile users can help themselves in self-screening, as a result of its small size and few parameters. The network should serve our life, so the lightweight network is very important. Research project risk management rules were used to propose these solutions. In the authors' opinion, however, the chances of having to use them are low.

Ethics approval and consent to participate

Ethics approval was obtained from The Second Hospital of Changzhou Affiliated to Nanjing Medical University. The study was in Accordance with guidelines mentioned. Medical ethics committee of The Second Hospital of Changzhou Affiliated to Nanjing Medical University waived the need for informed consent due to retrospective nature of the study.

Availability of data and materials

The datasets generated and/or analysed during the current study are not publicly available due the datasets are under continuous development and refinement, but are available from the corresponding author on reasonable request.

Consent for publication

Not applicable.

Competing interests

All the authors declare that there is no conflict of interest about the publication of this article.

Funding

This research was not supported by any grant funding.

Authors’ contribution

Guarantors of integrity of entire study, Mingzhu Meng and Long Wang; Data acquisition or data analysis/interpretation, Guangyuan He and Dong Shen; Manuscript drafting or manuscript revision for important intellectual content, all authors; Approval of final version of submitted manuscript, all authors; Literature research and clinical studies, Mingzhu Meng and Ming Zhang.

Acknowledgements

The authors wish to thank Lu Jing for his technical assistance in operating the Python programming code.

Authors’ information:¹ Mingzhu Meng Department of Radiology, The Second Hospital of Changzhou Affiliated to Nanjing Medical University, Changzhou, 213164, Jiangsu Province, P.R. China. ^*Long Wang Corresponding author: Email:[email protected]. Department of Radiology, The Second Hospital of Changzhou Affiliated to Nanjing Medical University, Changzhou, 213164, Jiangsu Province, P.R. China. Mingzhu Meng and Ming Zhang have the same contribution.

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(3):209-249.
Fujioka T, Mori M, Kubota K, Oyama J, Yamaga E, Yashima Y, Katsuta L, Nomura K, Nara M, Oda G et al: The Utility of Deep Learning in Breast Ultrasonic Imaging: A Review. Diagnostics (Basel) 2020, 10(12).
Shah SM, Khan RA, Arif S, Sajid U: Artificial intelligence for breast cancer analysis: Trends & directions. Comput Biol Med 2022, 142:105221.
Zerouaoui H, Idri A: Reviewing Machine Learning and Image Processing Based Decision-Making Systems for Breast Cancer Imaging. J Med Syst 2021, 45(1):8.
Gubern-Mérida A, Martí R, Melendez J, Hauth JL, Mann RM, Karssemeijer N, Platel B: Automated localization of breast cancer in DCE-MRI. Medical image analysis 2015, 20(1):265-274.
Zhang J, Xie Y, Wu Q, Xia Y: Medical image classification using synergic deep learning. Medical image analysis 2019, 54.:10-19.
Ayana G, Park J, Jeong JW, Choe SW: A Novel Multistage Transfer Learning for Ultrasound Breast Cancer Image Classification. Diagnostics (Basel) 2022, 12(1).
Saffari N, Rashwan HA, Abdel-Nasser M, Kumar Singh V, Arenas M, Mangina E, Herrera B, Puig D: Fully Automated Breast Density Segmentation and Classification Using Deep Learning. Diagnostics (Basel) 2020, 10(11).
Jiang Y, Edwards AV, Newstead GM: Artificial Intelligence Applied to Breast MRI for Improved Diagnosis. Radiology 2021, 298(1):38-46.
Das A, Nair MS, Peter SD: Computer-Aided Histopathological Image Analysis Techniques for Automated Nuclear Atypia Scoring of Breast Cancer: a Review. J Digit Imaging 2020, 33(5):1091-1121.
Dong F, She R, Cui C, Shi S, Hu X, Zeng J, Wu H, Xu J, Zhang Y: One step further into the blackbox: a pilot study of how to build more confidence around an AI-based decision system of breast nodule assessment in 2D ultrasound. European radiology 2021, 31(7):4991-5000.
Yang Y, Hu Y, Shen S, Jiang X, Gu R, Wang H, Liu F, Mei J, Liang J, Jia H et al: A new nomogram for predicting the malignant diagnosis of Breast Imaging Reporting and Data System (BI-RADS) ultrasonography category 4A lesions in women with dense breast tissue in the diagnostic setting. Quant Imaging Med Surg 2021, 11(7):3005-3017.
Tripathi S, Singh SK, Lee HK: An end-to-end breast tumour classification model using context-based patch modelling - A BiLSTM approach for image classification. Comput Med Imaging Graph 2021, 87:101838.
Barsha NA, Rahman A, Mahdy MRC: Automated detection and grading of Invasive Ductal Carcinoma breast cancer using ensemble of deep learning models. Comput Biol Med 2021, 139:104931.
Avanzo M, Wei L, Stancanello J, Vallieres M, Rao A, Morin O, Mattonen SA, El Naqa I: Machine and deep learning methods for radiomics. Med Phys 2020, 47(5):e185-e202.
Chan HP, Hadjiiski LM, Samala RK: Computer-aided diagnosis in the era of deep learning. Med Phys 2020, 47(5):e218-e227.
Litjens G, Kooi T, Bejnordi BE, Setio AAA, Ciompi F, Ghafoorian M, Laak JAWMvd, Ginneken Bv, Sánchez CI: A survey on deep learning in medical image analysis. Medical image analysis 2017, 42:60-88.
Sandler M, Howard A, Zhu M, Zhmoginov A, Chen L-C: MobileNetV2: Inverted Residuals and Linear Bottlenecks. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018，pp 4510-4520 2018.
G.Howard A, Zhu M, Chen B, Kalenichenko D, Wang W, Weyand T, Andreetto M, Adam H: MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications. arXiv:170404861[csCV] 2017:https://doi.org/10.48550/arXiv.41704.04861.
Wang Z, Li X, Yao M, Li J, Jiang Q, Yan B: A new detection model of microaneurysms based on improved FC-DenseNet. Sci Rep 2022, 12(1):950.
Mahmood T, Li J, Pei Y, Akhtar F: An Automated In-Depth Feature Learning Algorithm for Breast Abnormality Prognosis and Robust Characterization from Mammography Images Using Deep Transfer Learning. Biology (Basel) 2021, 10(9).
A Comparative Study of Pre-trained Convolutional Neural Networks for Semantic Segmentation of Breast Tumors in Ultrasound.
Wang W, Hu Y, Zou T, Liu[ H, Wang J, Wang X: A New Image Classification Approach via Improved MobileNet Models with Local Receptive Field Expansion in Shallow Layers. Comput Intell Neurosci 2020, 2020:8817849.
Arora V, Ng EY, Leekha RS, Darshan M, Singh A: Transfer learning-based approach for detecting COVID-19 ailment in lung CT scan. Comput Biol Med 2021, 135:104575.
Taresh MM, Zhu N, Ali TAA, Hameed AS, Mutar ML: Transfer Learning to Detect COVID-19 Automatically from X-Ray Images Using Convolutional Neural Networks. Int J Biomed Imaging 2021, 2021:8828404.

Tables 1 to 6 are available in the Supplementary Files section.

No competing interests reported.

Download PDF

Version 2

posted

You are reading this latest preprint version

Classification of breast lesions on DCE-MRI Data Using Fine-tuned MobileNet

Status:

Version 2

Abstract

Background

Methods

Results

Conclusion

Figures

Introduction

Materials And Methods

Results

Discussion

Conclusion

Declarations

References

Tables

Additional Declarations

Supplementary Files

Status:

Version 2