Predicting Malignancy in Breast Lesions: Enhancing Accuracy with Fine-Tuned Convolutional Neural Network Models

doi:10.21203/rs.3.rs-3937557/v1

Download PDF

Research Article

Predicting Malignancy in Breast Lesions: Enhancing Accuracy with Fine-Tuned Convolutional Neural Network Models

https://doi.org/10.21203/rs.3.rs-3937557/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Objectives.

This study aimed to explore which convolutional neural network (CNN) model is best for predicting the likelihood of malignancy on dynamic contrast-enhanced breast magnetic resonance imaging (DCE-BMRI).

Materials and Methods.

A total of 273 benign (benign group) and 274 malignant lesions (malignant group) were obtained, and randomly divided into a training set (benign group: 246 lesions, malignant group: 245 lesions) and a testing set (benign group: 28 lesions, malignant group: 28 lesions) in a 9:1 ratio. An additional 53 lesions from 53 patients were designated as the validation set. Five models (VGG16, VGG19, DenseNet201, ResNet50, and MobileNetV2) were evaluated. The metrics for model performance evaluation included accuracy (Ac) in the training and testing sets, and precision (Pr), recall rate (Rc), F1 score (F1), and area under the receiver operating characteristic curve (AUC) in the validation set.

Results.

Accuracies of 1.0 were achieved on the training set by all five fine-tuned models (S1-5), with model S4 demonstrating the highest test accuracy at 0.97. Additionally, S4 showed the lowest loss value in the testing set. The S4 model also attained the highest AUC (Area Under the Curve) of 0.89 in the validation set, marking a 13% improvement over the VGG19 model. Notably, the AUC of S4 for BI-RADS 3 was 0.90 and for BI-RADS 4 was 0.86, both significantly higher than the 0.65 AUC for BI-RADS 5.

Conclusion.

The S4 model we propose emerged as the superior model for predicting the likelihood of malignancy in DCE-BMRI and holds potential for clinical application in patients with breast diseases. However, further validation is necessary, underscoring the need for additional data.

BI-RADS

Convolutional Neural Networks

Deep transfer learning

Breast lesions

Magnetic resonance imaging

In 2020, breast cancer accounted for 2.3 million new cases among women, surpassing lung cancer as the most commonly diagnosed cancer in this demographic, with a prevalence of 11.7%[1].Despite a continued decline in breast cancer mortality in the United States, which saw a 40% reduction between 1989 and 2017, there was a notable 0.3% annual increase in incidence rates over a five-year period (2012–2016). This rise is primarily attributed to increasing rates of local stage and hormone receptor-positive diseases[2]. While China’s cancer incidence rate remains lower than those of the United Kingdom and the United States, the expected rise in cancer cases in the coming years is a concern. This anticipated increase is due to an aging population, population growth, and more prevalent westernized lifestyles[3]. Consequently, there has been a surge in female breast cancer cases in China, mirroring a trend also observed in developed countries like the USA and UK [4]. Recently, the diagnosis of breast cancer has become increasingly complex, owing to a more comprehensive understanding of the hallmark characteristics of breast tumors[5, 6]. Early detection and accurate diagnosis are imperative for effective treatment and improved outcomes, ultimately contributing to a reduction in the breast cancer mortality rate.

The Breast Imaging Reporting and Data System (BI-RADS) prototype, first published by the American College of Radiology (ACR) in 1993, addressed the lack of uniformity in mammography reporting and has since undergone several revisions[7–9]. BI-RADS has gained wide acceptance among clinicians and radiologists.

The fifth edition, updated in 2013, further clarified the professional terminology for breast cancer and developed standardized imaging reports. Efforts were made to ensure compatibility across all three imaging lexicons ((mammography, ultrasound, and magnetic resonance imaging (MRI)) by uniformly referring to lesions.

BI-RADS category 3 (probably benign) findings warrant short-term follow-up, while category 4 and 5 lesions require histopathological examination for definitive conclusions. Most BI-RADS 3, 4, and 5 lesions are diagnosed as benign. Besides the additional costs incurred from further examinations and procedures, patients also experience increased psychological stress, anxiety, and inconvenience.

Consequently, reclassifying more benign lesions as BI-RADS category 1 or 2 could create a win-win situation for both patients and health systems. Dynamic contrast-enhanced breast magnetic resonance imaging (DCE-BMRI) may offer better differentiation between malignant and benign features, potentially reducing unnecessary imaging follow-ups and benign biopsies, although current literature lacks consistency in this regard.

Presently, various methods are employed for early detection of breast cancer. While many screening methods are in use, they each have limitations. In this study, we utilized a state-of-the-art convolutional neural network (CNN) to classify breast lesions, assessing its potential to improve diagnostic accuracy. Deep transfer learning (DTL) techniques have been successfully applied in medical image analysis[10–13], with pre-trained neural networks being used to differentiate between benign and malignant breast lesions in ultrasound images[14]. However, the application of these techniques to DCE-BMRI images has been limited. Therefore, the primary aim of this study was to determine which pre-trained model most effectively predicts the likelihood of malignancy in DCE-BMRI.

2.1 Dataset 1: Training and Testing Set

We collected data from 530 patients with complete DCE-BMRI and pathological information, spanning January 2017 to December 2020. This included 17 patients with bilateral lesions (both benign and malignant lesions on one side). All lesions were pathologically confirmed and categorized into benign or malignant groups. These were then randomly assigned to a training set (benign: 246 lesions, malignant: 245 lesions) and a testing set (benign: 28 lesions, malignant: 28 lesions) in a 9:1 ratio (refer to Fig. 1). Variables such as age, pathological type, and tumor diameter were compared between groups. Table 1 details the pathological distribution of breast lesions. Inclusion criteria were: ① Patients not subjected to preoperative chemotherapy or chemoradiotherapy before MRI, ② Absence of puncture or surgical procedures prior to MRI. Due to space constraints, clinical presentation details are omitted. To minimize bias from bilateral lesions, only unilateral DCE-BMRI images were used.

Table 1

The pathological distribution of breast lesions
Pathological diagnosis	Lesions	Percent (%)
Malignant lesions
Invasive ductal carcinoma	220	80.29
Intraductal carcinoma	33	12.04
Invasive lobular carcinoma	7	2.55
Mucinous carcinoma	10	3.65
Lymphoma	1	0.36
Papillary carcinoma	3	1.09
Total	274	100.00
Benign lesions
Cyst	26	9.52
Adenosis	42	15.38
Fibroadenoma	176	64.47
Chronic inflammation	6	2.20
Intraductal papilloma	20	7.33
Lobular tumor	3	1.10
Total	273	100.00

2.2 Dataset 2: Validation Set

Simultaneously, 53 lesions from 53 patients were included as Dataset 2, using the same MRI scanner as Dataset 1, but unseen during training. Dataset 2 comprised three subsections: BI-RADS 3, 4, and 5 (see Fig. 1). Surgical pathology confirmed final diagnoses in cases where percutaneous biopsy indicated high risk. Absence of surgery with imaging stability was deemed indicative of no associated cancer. Follow-up adhered to referenced criteria[9, 15]. Correct classification of a lesion required accurate classification in six out of ten images. Table 2 lists the specific details of Dataset 2.

Table 2

Patients in dataset 2
Category	N		Confirmed		follow up
Category	B	M	pathologically		follow up
BI-RADS 3	10	10		4	16
BI-RADS 4	10	10		14	6
BI-RADS 5	3	10		13	0

2.3 MRI Techniques

We employed two 3T MRI scanners with dedicated breast coils in a prone position. Gd-DTPA (0.1 mmol/kg, 2.50 mL/s) was injected through the elbow vein. The process involved six dynamic enhancement phases (one pre-contrast, five post-contrast). MRIs were conducted preoperatively and before initiating therapy. Detailed scanning parameters are outlined in Table 3.

Table 3

Scan parameters for the two magnetic resonance scanners
Parameter	Philips Achieva	GE Healthcare
Field strength	3.0T	3.0T
No. of coil channels	8	8
Acquisition plane	Axial	Axial
Pulse sequence	3D gradient echo (Thrive)	Enhanced fast gradient echo 3D
Repetition time (ms)	5.5	9.6
Echo time (ms)	2.7	2.1
Flip angle	10°	10°
No. of postcontrast sequence	5	5
Fat suppression	Yes	Yes
Scan time	570s	500s
3D, three dimensional; ms, millisecond; s: second.

2.4 Readers

Five experienced radiologists from our department, each with over five years of breast MRI interpretation experience and specialized training in breast imaging, were enlisted. MRI image analyses were conducted using the GOLDPACS viewer (www.jinpacs.com).

2.5 Proposed model

The study utilized a computer equipped with an Intel (R) Core (TM) i7-10700F, NVIDIA RTX 2060 GPU, running on Windows 10 Enterprise 64-bit with 6GB RAM. All extraneous programs were closed during model operation. Each network underwent identical data testing and training for consistent comparison. Malignant images were identified based on a threshold of ≥ 0.5, while images below this threshold were considered benign.

We selected five commonly used pretrained models (VGG16, VGG19, DenseNet201, ResNet50, and MobileNetV2) and employed five-fold cross-validation to assess model performance, selecting the best-performing model. This cross-validation process was then applied to Dataset 2. Additionally, we enhanced model performance using various fine-tuning strategies. The architecture of the proposed DTL with the five models for breast lesion classification is depicted in Fig. 2.

Initially, the images underwent random shuffling. Data augmentation techniques (rotation, shear range, zoom range, and horizontal flip) were applied prior to training. The binary cross-entropy loss function was used, and the training process was optimized using the Adam optimizer with a learning rate of 0.001. Our model required 200 epochs for training on DCE-BMRI images, with a batch size of 64 images. Activation functions included ReLU and sigmoid, as detailed in Equations 1 and 2

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

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

2.6 Evaluation metrics

We assessed the effectiveness of Deep Transfer Learning (DTL) models using five performance metrics: accuracy (Ac), precision (Pr), recall rate (Rc), F1 score (F1), and the area under the receiver operating characteristic curve (AUROC)[16]. For this analysis, cases were classified as either malignant or benign, representing positive and negative cases, respectively. True positives (TP) and true negatives (TN) denote the proportion of correctly diagnosed malignant and benign cases. False positives (FP) and false negatives (FN) indicate lesions misdiagnosed as benign and malignant, respectively. The formulas for these metrics are as follows:

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

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

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

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

Notably, the accuracy metric (Ac) does not account for data distribution. The F1 score is a balanced measure that considers both precision and recall, making it particularly useful in datasets with imbalanced classes.

2.7 Statistical analysis

Statistical analyses were conducted using SPSS 23.0 software (IBM). For data adhering to a normal distribution, counting data were presented as mean ± standard deviation ($\stackrel{-}{\text{x}}$ ± s). One-way analysis of variance (ANOVA) was employed for variance analysis between groups. The Mann-Whitney U test was applied for data not meeting the normal distribution criteria. The chi-square test was utilized for comparing frequency counts between malignant and benign groups in the datasets (training and testing sets). A P-value of < 0.05 (two-tailed) was considered statistically significant.

3.1 Age and Lesion Diameter

Age and lesion diameter did not conform to normal distribution. The age difference between the malignant group (46.40 ± 10.90 years) and the benign group (44.84 ± 10.20 years) was not statistically significant (P = 0.136). However, lesion diameters were significantly smaller in the malignant group (25.06 ± 11.54 mm) compared to the benign group (33.44 ± 16.69 mm) (P < 0.001). No significant variance was observed in lesion distribution between the training and testing sets across both groups (P = 0.988).

3.2 Cross validation

We evaluated five models (VGG16, VGG19, DenseNet201, ResNet50, and MobileNetV2) through five-fold cross-validation in Dataset 1 (see Table 4 for results). The DenseNet201 and MobileNetV2 models achieved perfect accuracy (1.00) in the training set, but their testing set accuracies were lower at 0.91 and 0.88, respectively, both below VGG19’s 0.96. Despite similar architectures, VGG19 outperformed VGG16 (0.91). However, both VGG16 and VGG19 exhibited premature loss increases with epoch advancement, indicating non-convergence on Dataset 1 and potential overfitting. Similar trends were observed for MobileNetV2 and DenseNet201. ResNet50 showed the lowest accuracy among the models (0.92 training, 0.67 testing). Figures 3 and 4 illustrate the learning curves and heat maps.

Table 4

The results of the five-fold cross-validation
Folds	Accuracies of the training set					Accuracies of the testing set
Folds	model1	model2	model3	model4	model5	model1	model2	model3	model4	model5
Fold1	1.00	1.00	1.00	0.92	1.00	0.91	0.96	0.91	0.67	0.88
Fold2	1.00	1.00	1.00	0.93	0.99	0.90	0.96	0.91	0.67	0.87
Fold3	1.00	1.00	1.00	0.92	1.00	0.91	0.96	0.91	0.67	0.88
Fold4	1.00	1.00	1.00	0.92	0.99	0.90	0.96	0.91	0.67	0.87
Fold5	1.00	1.00	1.00	0.93	1.00	0.91	0.96	0.91	0.67	0.88

Note. Model1, VGG16; Model2, VGG19; Model3, DenseNet201; Model4, ResNet50; Model5, MobileNetV2.

3.3 Fine-tuning strategy

Given these findings, we focused on enhancing the VGG19 model through five distinct fine-tuning strategies (Fig. 5). The fine-tuning involved activating neural network parameters for training, while keeping certain layers frozen. We noted that the accuracy achieved was 1.0, for all five fine-tuning models(S1-5) on the training set, but S4 obtained the highest test accuracy of 0.97 on the testing set. In addition, the loss value was the lowest in the testing set for S4. These results reveal that the S4 model has a better generalization ability than the other fine-tuned models.

3.4 ROC Analysis on Validation Set

Analysis of the Receiver Operating Characteristic (ROC) curve for the five models on the validation set revealed VGG19 as the highest performer (AUC 0.92), yet the validation set AUC was only 0.76. Among the fine-tuned models, S4 attained the highest AUC (0.89) on the validation set, marking a 13% improvement over the original VGG19 (Fig. 6). Further analysis of S4 across BI-RADS categories 3, 4, and 5 showed notably higher AUCs for BI-RADS 3 (0.90) and 4 (0.86) compared to 5 (0.65) (Fig. 7).

3.5 Classification Reports on Validation Set

Classification reports for the five models and S1-5 strategies are provided in Table 5. For the validation set, VGG19 achieved higher performance metrics (Pr 0.75, Rc 0.76, F1 0.73, AUC 0.76) compared to the other models. Strategy S4 outperformed all others on the validation set with Pr 0.89, Rc 0.88, F1 0.87, and AUC 0.89.

Table 5

Classification report of deep transfer learning models in validation set
DTL models		Pr			Rc			$\text{f}1$			AUC
				group1	group2	avg	group1	group2	avg	group1	group1	avg
		model1	0.93		0.53	0.73	0.60	0.91	0.75	0.73	0.67	0.70	0.76
model2	0.91		0.59	0.75	0.63	0.90	0.76	0.75	0.71	0.73	0.73
model3	0.91		0.52	0.71	0.59	0.88	0.74	0.72	0.65	0.69	0.71
model4	0.90		0.39	0.65	0.53	0.84	0.68	0.67	0.53	0.60	0.65
model5	0.91		0.55	0.73	0.61	0.89	0.75	0.73	0.68	0.71	0.73
S1	0.88		0.63	0.75	0.65	0.87	0.76	0.74	0.73	0.74	0.75
S2	0.95		0.60	0.77	0.64	0.94	0.79	0.77	0.73	0.75	0.77
S3	0.91		0.60	0.76	0.63	0.91	0.76	0.74	0.71	0.73	0.75
S4	0.98		0.79	0.89	0.78	0.98	0.88	0.87	0.88	0.87	0.89
S5	0.94		0.59	0.77	0.64	0.93	0.79	0.76	0.73	0.75	0.77

Note. Group1, benign group; Group2, malignant group; Avg, average; Model1, VGG16; Model2, VGG19; Model3, DenseNet201; Model4, ResNet50; Model5, MobileNetV2.

In this study, we evaluated five pre-trained convolutional neural network models using a 5-fold cross-validation method on our DCE-BMRI dataset. Our objective was to identify the best-performing model, which we defined as the one excelling across all predefined evaluation criteria. Following this, we fine-tuned the chosen model to enhance its performance further and tested its generalization capability on a validation set.

Our findings indicated that the VGG19 model demonstrated superior performance, achieving accuracies of 1.00 and 0.96 on the training and testing sets, respectively. Moreover, VGG19 achieved the highest Area Under the Curve (AUC) of 0.92 on the first validation set, but this dropped to 0.76 on a subsequent validation set, suggesting limitations in its generalization ability. Previous research supports the notion that fine-tuning can enhance the accuracy and precision of such models[17–19].

Consequently, we developed five distinct fine-tuning strategies for VGG19. Strategy S4 emerged as the most successful, yielding the highest test accuracy (0.97) and the lowest test loss on the validation set, indicating a superior generalization capability compared to the other strategies. When comparing the AUC scores of strategies S1-5 on the validation set, S4 again scored highest with an AUC of 0.89. These results are promising for advancing the accuracy of medical image classification diagnostics.

We also delved into whether the S4 model exhibited different AUC scores across BI-RADS categories 3, 4, and 5. Interestingly, S4 performed best in BI-RADS 3 (AUC 0.90), followed by BI-RADS 4 (AUC 0.86), and showed the least performance in BI-RADS 5 (AUC 0.65). The BI-RADS 3 category, typically applied when the likelihood of cancer is less than 2%, aims to minimize unnecessary biopsies for pathologically benign findings. However, patient compliance with follow-up MRI recommendations every six months is notably low in this category[20].

The BI-RADS system is set to evolve with new breast imaging modalities. Key areas for improvement include expanding the lexicon for common findings and clarifying the application of Category 3[21]. BI-RADS 3 represents a significant portion (13.9%) of diagnostic exams [22], often leading to follow-up procedures for patients classified as ‘probably benign’, yet compliance with these follow-up recommendations remains a challenge [20]. This lack of compliance raises concerns about the clinical and economic implications, particularly regarding the resolution time and outcomes for patients in this category.

In a particular study, only 1.4% of BI-RADS 3 lesions were found to be malignant, including two cases of delayed diagnosis at 13.2 and 33.2 months, respectively. The incidence of delayed diagnosis due to additional MRI-detected lesions during follow-up was notably low (0.7%), consisting exclusively of T1N0 contralateral cancers. This finding suggests that annual follow-up may suffice for BI-RADS 3 lesions identified by MRI before surgery[15]. Consequently, accurately distinguishing between benign and malignant lesions in BI-RADS 3 is crucial. Our study potentially offers significant benefits to a substantial number of patients diagnosed with BI-RADS category lesions using DCE-BMRI imaging.

BI-RADS category 4 lesions are associated with a high likelihood of malignancy, with estimates ranging from 2–95%. The BI-RADS 4 classification, to a degree, is subjective; the outcomes of biopsies in this category vary significantly, and the rate of cancer detection relative to the number of biopsies performed is relatively low (17.8%)[15]. Moreover, unnecessary biopsies can result in a range of adverse effects, including pain, fear, emotional distress, and financial costs.

Breast MRI is highly sensitive, yet it often presents a challenge in differentiating between atypical malignant and benign lesions, leading to potential overclassification in the BI-RADS 4 category and subsequent invasive biopsies. The wide range of positive predictive values for MRI-guided biopsies (2.5–84.0%) [23, 24], indicates that many patients undergo unnecessary procedures. Indeed, numerous women subjected to biopsies for benign findings endure unnecessary discomfort, expenses, potential complications, cosmetic alterations, and anxiety [25]. Identifying predictors of benign BI-RADS 4 masses, therefore, could be highly beneficial.

Efforts to enhance the assessment of BI-RADS 4 lesions could improve the identification of benign lesions, thereby reducing the frequency of unnecessary biopsies. Some researchers have developed predictive models based on imaging features or multiparameter MRI data to better evaluate BI-RADS 4 lesions, though these models typically rely on traditional imaging features subjectively defined by radiologists[26].

The BI-RADS 5 category is applied when imaging findings suggest a malignancy probability of 95% or higher. According to MC et al.[27], the positive predictive value of BI-RADS 5 assessments is only 71.4%, indicating that not all lesions classified as BI-RADS 5 are malignant[28], and surgery is often recommended for this category. It is recognized that a single imaging finding rarely confers such a high risk of malignancy; rather, a combination of features is necessary to elevate a lesion to Category 5[7]. However, it's important to acknowledge that even when tissue samples from molecular biopsies are used, they may not fully represent the entire lesion, as biopsies often target only a small, specific area of a heterogeneous lesion, introducing a bias in lesion selection.

Rc also known as sensitivity, measures a classifier's completeness. A lower Rc value indicates the classifier's limited capability in handling large FP values. Recent publications have led to the introduction of new and updated performance benchmarks in the latest edition, replacing outdated metrics. The recall rate benchmark has been revised accordingly. Initially, about half of all radiologists were unable to meet the 10% benchmark for recall rate, prompting a revision to a more achievable target of 12%, a standard met by over 75% of radiologists[7]. Our study showed that the S4 model exhibited the highest recall rate (0.89) among all DTL models, which is notable given the relatively limited class diversity in our dataset.

This study, however, is not without limitations. Firstly, the training set contained a relatively small number of images, particularly lacking in rare lesion types. Consequently, our dataset may not fully represent the broader spectrum of breast disease patients, potentially impacting the accuracy of the DTL model. Therefore, it’s essential to conduct further analysis with more extensive datasets to comprehensively evaluate the robustness of the DTL model. Secondly, our study relied solely on static Dynamic Contrast-Enhanced Breast Magnetic Resonance Imaging (DCE-BMRI) images, excluding other routine diagnostic procedures such as clinical evaluations, breast ultrasounds, and mammography. Thirdly, we limited our investigation to just five pre-trained models; future research should explore a wider range of models to determine their robustness on larger datasets. Lastly, while this paper does not delve into the various methods of fine-tuning Convolutional Neural Network (CNN) models, these topics will be the focus of our subsequent studies.

The S4 model demonstrated superior accuracy in BI-RADS categories 3 and 4 compared to category 5. This finding is significant as it could lead to a reduction in the number of follow-up sessions for BI-RADS 3 and decrease the number of unnecessary biopsies for benign lesions in BI-RADS 4. However, these results necessitate further validation. Moving forward, our focus will be on exploring more robust models and the importance of augmenting our dataset with additional data.

MRI= Magnetic resonance imaging

DL= Deep learning

DTL=Deep transfer learning

ROC = receiver operating characteristic

AUC = area under the ROC curve

BI-RADS = Breast Imaging Reporting and Data System,

DCE-BMRI = dynamic contrast enhanced breast MRI

Ethics approval and consent to participate

This study was approved by the Second Hospital of Changzhou Affiliated to Nanjing Medical University of Chinese Medicine Ethics Review Committee (Ethics Number: [2023]KY313-01).

Consent for publication

This study was a retrospective analysis and informed consent was waived.

Availability of data and materials

The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.

Competing interests

The authors declare that they have no competing interests.

Funding

This study was supported by the Program of Bureau of Science and Technology Foundation of Changzhou (No. CJ20220260).

Authors’ contributions

Li Li and Mingzhu Meng carried out the literature search, and designed and wrote the manuscript. Mingzhu Meng and Changjie Pan conceived of the project, and participated in its design and coordination, and helped to draft the manuscript. Ming Zhang, Dong Shen and Guangyuan He are responsible for figures processing. Both authors read and approved the final manuscript.

Acknowledgments

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

Sung H, Ferlay J, Siegel RL,et al. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin. 2021;71(3):209-49.
DeSantis CE, Ma J, Gaudet MM,et al. Breast cancer statistics, 2019. CA Cancer J Clin. 2019;69(6):438-51.
Cao W, Chen HD, Yu YW,et al. Changing profiles of cancer burden worldwide and in China: a secondary analysis of the global cancer statistics 2020. Chin Med J (Engl). 2021;134(7):783-91.
RM F, YN Z, SM C,et al. Current cancer situation in China: good or bad news from the 2018 Global Cancer Statistics? Cancer Commun (Lond). 2019;39(1):22.
Gao Y, Heller SL. Abbreviated and Ultrafast Breast MRI in Clinical Practice. RadioGraphics. 2020;40(6):1507-27.
Berdzuli N. Breast cancer: from awareness to access. BMJ. 2023;380(290.
Mercado CL. BI-RADS Update. Radiol Clin North Am. 2014;52(3):481-7.
Sedgwick EL, Ebuoma L, Hamame A,et al. BI-RADS update for breast cancer caregivers. Breast Cancer Research and Treatment. 2015;150(2):243-54.
Pesce K, Orruma MB, Hadad C,et al. BI-RADS Terminology for Mammography Reports: What Residents Need to Know. Radiographics. 2019;39(2):319-20.
Huang Z, Zhou Q, Zhu X,et al. Batch Similarity Based Triplet Loss Assembled into Light-Weighted Convolutional Neural Networks for Medical Image Classification. Sensors (Basel). 2021;21(3):764-85.
Tajbakhsh N, Jeyaseelan L, Li Q,et al. Embracing imperfect datasets: A review of deep learning solutions for medical image segmentation. Medical image analysis. 2020;63:101693.
Naqa IE. The role of machine and deep learning in modern medical physics. Med Phys. 2020;47(5):e125-e6.
Zhang J, Xie Y, Wu Q,et al. Medical image classification using synergic deep learning. Medical image analysis. 2019;54:10-9.
Dong F, She R, Cui C,et al. 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.
Lee SE, Lee JH, Han K,et al. BI-RADS category 3, 4, and 5 lesions identified at preoperative breast MRI in patients with breast cancer: implications for management. Eur Radiol Exp. 2020;30(5):2773-81.
Wang Z, Li X, Yao M,et al. A new detection model of microaneurysms based on improved FC-DenseNet. Sci Rep. 2022;12(1):950.
Tan T, Li Z, Liu H,et al. Optimize Transfer Learning for Lung Diseases in Bronchoscopy Using a New Concept: Sequential Fine-Tuning. IEEE J Transl Eng Health Med. 2018;6:1800808.
Ahamed KU, Islam M, Uddin A,et al. A deep learning approach using effective preprocessing techniques to detect COVID-19 from chest CT-scan and X-ray images. Comput Biol Med. 2021;139:105014.
Montaha S, Azam S, Rafid A,et al. BreastNet18: A High Accuracy Fine-Tuned VGG16 Model Evaluated Using Ablation Study for Diagnosing Breast Cancer from Enhanced Mammography Images. Biology (Basel). 2021;10(12):1347.
Ambinder EB, Myers K, Panigrahi B,et al. Breast MRI BI-RADS 3: Impact of Patient-Level Factors on Compliance With Short-Term Follow-Up. J Am Coll Radiol 2020;17(3):377-83.
Eghtedari M, Chong A, Rakow-Penner R,et al. Current Status and Future of BI-RADS in Multimodality Breast Imaging, From the AJR Special Series on Radiology Reporting and Data Systems. American Journal of Roentgenology. 2020:[published online].
Elezaby M, Li G, Bhargavan-Chatfield M,et al. ACR BI-RADS assessment category 4 subdivisions in Diagnostic Mammography: Utilization and Outcomes in the National Mammography Database. Radiology. 2018;287(2):416-22.
RM S, ES B, M E,et al. Utility of BI-RADS Assessment Category 4 Subdivisions for Screening Breast MRI. AJR American journal of roentgenology. 2017;208(6):1392-9.
JR MdA, AB G, TP B,et al. Subcategorization of Suspicious Breast Lesions (BI-RADS Category 4) According to MRI Criteria: Role of Dynamic Contrast-Enhanced and Diffusion-Weighted Imaging. AJR American journal of roentgenology. 2015;205(1):222-31.
Laws A, Crocker A, Dort J,et al. Improving Wait Times and Patient Experience Through Implementation of a Provincial Expedited Diagnostic Pathway for BI-RADS 5 Breast Lesions. Ann Surg Oncol 2019;36(10):3361-7.
Hao W, Gong J, Wang S,et al. Application of MRI Radiomics-Based Machine Learning Model to Improve Contralateral BI-RADS 4 Lesion Assessment. Frontiers in Oncology. 2020;10:531476-84.
MC M, C G, L H,et al. Positive Predictive Value of BI-RADS MR imaging. Radiology. 2012;264(1):51-8.
Dao KA, Rives AF, Quintana LM,et al. BI-RADS 5: More than Cancer. Radiographics. 2020;40(5):1203-4.

No competing interests reported.

DataS1.doc

Download PDF

Editorial decision: Revision requested
19 Aug, 2024
Reviews received at journal
26 Jul, 2024
Reviews received at journal
21 Jul, 2024
Reviewers agreed at journal
19 Jul, 2024
Reviewers agreed at journal
19 Jul, 2024
Reviewers invited by journal
19 Apr, 2024
Editor invited by journal
15 Feb, 2024
Submission checks completed at journal
15 Feb, 2024
Editor assigned by journal
15 Feb, 2024
First submitted to journal
07 Feb, 2024

You are reading this latest preprint version

Predicting Malignancy in Breast Lesions: Enhancing Accuracy with Fine-Tuned Convolutional Neural Network Models

Status:

Version 1

Abstract

Objectives.

Materials and Methods.

Results.

Conclusion.

Figures

1 Introduction

2 Materials and Methods

2.1 Dataset 1: Training and Testing Set

2.2 Dataset 2: Validation Set

2.3 MRI Techniques

2.4 Readers

2.5 Proposed model

2.6 Evaluation metrics

2.7 Statistical analysis

3 Results

3.1 Age and Lesion Diameter

3.2 Cross validation

3.3 Fine-tuning strategy

3.4 ROC Analysis on Validation Set

3.5 Classification Reports on Validation Set

4 Discussion

5 Conclusions

Abbreviations

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1