Combining shallow and deep neural networks on pseudo-color enhanced images for digital breast tomosynthesis lesion classification

Abstract

Introduction:

The classification of lesion types in Digital Breast Tomosynthesis (DBT) images is crucial for the early diagnosis of breast cancer. However, the task remains challenging due to the complexity of breast tissue and the subtle nature of lesions. To alleviate radiologists’ workload, computer-aided diagnosis (CAD) systems have been developed. The breast lesion regions vary in size and complexity, which leads to performance degradation.

Methods:

To tackle this problem, we propose a novel DBT Dual-Net architecture comprising two complementary neural network branches that extract both low-level and high-level features. By fusing different-level feature representations, the model can better capture subtle structure. Furthermore, we introduced a pseudo-color enhancement procedure to improve the visibility of lesions on DBT. Moreover, most existing DBT classification studies rely on two-dimensional (2D) slice-level analysis, neglecting the rich three-dimensional (3D) spatial context within DBT volumes. To address this limitation, we used majority voting for image-level classification from predictions across slices.

Results:

We evaluated our method on a public DBT dataset and compared its performance with several existing classification approaches. The results showed that our method outperforms baseline models.

Discussion:

The use of pseudo-color enhancement, extracting high and low-level features and inter-slice majority voting proposed method is effective for lesion classification in DBT. The code is available at https://github.com/xiaoerlaigeid/DBT-Dual-Net.

1 Introduction

Breast cancer remains one of the life-threatening diseases in women (1). Among the diagnostic imaging modalities, mammography has long been established as the standard screening method. Despite its widespread use, traditional mammography has obvious limitations, particularly in dense breast tissue, where overlapping structures can obscure lesions or lead to false positives (2). Digital breast tomosynthesis (DBT) is an imaging method that reconstructs a 3-dimensional (3D) volume of the entire breast using a series of low-dose 2D X-ray projections captured from various angles. It retains spatial information across the Z direction, improving the detection of tumors in dense tissues (3). Furthermore, DBT improves detailed visualization of breast structures, which could reduce the false positive rate. However, by producing 3D image data, it may increase the radiologists’ workload.

With the development of computer-aided diagnosis (CAD) techniques, many radiomics have been applied in medical diagnosis, demonstrating potential for classifying breast tumors in DBT images (4–6). For example, (7) used linear regression analysis to explore the relationship between radiomics features in DBT images and tumors. (8) evaluated multiple machine learning methods for DBT tumor classification. Their approaches also utilized radiomics features to enable comprehensive image analysis. Several studies explored the relationship between radiomics features from DBT images and the Ki-67 expression level to predict tumor characteristics (9–11). These features may include handcrafted features like morphological, gray-level statistics, and texture features, or contain features extracted with deep learning.

The analysis of DBT images using radiomics techniques remain time-consuming, as it require radiologists to manually delineate suspicious regions. Recently, the application of deep learning in CAD for DBT has gained traction. Numerous studies indicate that deep learning holds significant promise for the classification of breast lesions (12–21).

To tackle the complexity of breast lesions, several studies adopted various approaches to pre-process the dataset. Some studies used separate slices as inputs and report performance at the slice level (13, 15, 16, 18, 19, 22). Kassis et al. (16) employed a vision transformer (ViT) to extract features from 2D DBT slices, followed by post-processing that incorporated information from adjacent slices to classify the entire 3D image. Zhang et al. (21) proposed a novel architecture that includes early fusion and late fusion on multiple-slice 2D DBT images. Early fusion refers to averaging multiple slice images to create a new image, while late fusion involves using a feature extractor to extract features from multiple slices, and then applying a pooling layer to obtain a single feature map. Their study made use of all the information from an individual patient’s lesion case. To fully leverage 3D volumetric information while addressing the challenge of variable slice counts across cases, (17) proposed a preprocessing approach that utilizes a fixed-size (a subset of sections) to replace the complete 3D volume by incorporating a ViT-based network model.

However, the whole-image methods lack interpretability because they do not show a clear causal region to support the prediction. Therefore, some studies directly used region of interest (ROI) lesion patches as inputs. Some researchers have adopted pre-trained results from mammogram ROI data for DBT patch classification tasks using transfer learning (23–25). Samala et al. (24) compared the performance of two approaches with fine-tuning: one was with a multi-stage model, where the network was first fine-tuned on mammograms after pretraining on ImageNet and then further fine-tuned on DBT data, whereas the other was with a model pretrained on ImageNet and directly fine-tuned on DBT patches. Other studies fused information from different ROI patches within the same case. In Wang et al. (26), three multifaceted representations of the breast mass (gross mass, overview, and mass background) were extracted from the ROI and independently processed by a multi-scale multi-level features-enforced DenseNet.

Previous DBT classification work has primarily focused on classifying individual slices, which may not accurately represent the entire image-level structure of breast lesions. Some works attempt to crop whole images into different sections and sample neighboring sections as inputs (15, 17, 27). Predictions from different slices of the same case may yield conflicting results when processed independently by the model. Additionally, the breast lesion width varied in size from 10 mm to even more than 50 mm, which will harm the deep learning classification performance (28).

Another trend in recent studies is to focus on deeper neural network architectures to extract complex hierarchical features, while neglecting potentially valuable information captured by shallower neural network models (29). Some studies have compared the performance of deep and shallow networks in other breast image diagnosis systems. Kaya et al. (30) investigated this through an internal classifier network design, concluding from experimental results that deeper networks may induce overthinking phenomena that lead to computational resource waste that can be destructive for prediction tasks. The choice between these architectures is currently done by trial and error (31). Consequently, numerous experiments have attempted to incorporate shallow feature branches into deeper network architectures, which explicitly model and collaboratively optimize both shallow and deep features for tasks such as image classification and segmentation. For example, (32) proposed a Dual-Branch-UNet model for the vascular segmentation task, in which the deeper branches input low-resolution images and the shallower network branches process high-resolution images to extract spatial detail information. Gao et al. (33) proposed a novel Shallow-Deep CNN architecture for breast cancer diagnosis. In this architecture, a shallow network synthesizes contrast-enhanced-like images from standard mammograms, while a deep network analyzes lesion features in actual contrast-enhanced scans. These methods are inspired by the procedure performed by radiologists in which they combine different types of features (e.g., morphology and appearance) to perform diagnosis (34), which can potentially be extracted with different subnetworks.

Many medical imaging modalities only provide grayscale images, yet numerous studies have demonstrated that incorporating color information through colorization methods can significantly enhance various downstream tasks. These approaches operate under the hypothesis that pseudo-color enhancement algorithms (PCE), which transform grayscale images into pseudo-color representations, improve visual clarity and accentuate subtle details for both human interpretation and computational models, a premise substantiated by experimental results across multiple domains. El-Shazli et al. (15) applied PCE to the DBT grayscale images; adjustments in the hue, saturation, and value (HSV) color space, particularly the hue channel adjustments, were helpful to differentiate between different tissue regions. This approach has been assumed to avoid the coupling effect of simultaneously modifying color and brightness in RGB channels.

Additionally, Huang et al. (35) achieved superior segmentation accuracy in ultrasound tumor detection using PCE transformations. Zhang et al. (36) developed a specialized colorization network model that generates anatomically vivid tumor representations from PET-CT scans. Beyond conventional intensity-based color mapping, researchers like (37, 38) have pioneered multi-slice fusion techniques, where they composite different slice views of breast lesions into three-channel pseudo-RGB images, subsequently demonstrating measurable improvements in breast mass classification performance.

Inspired by the shallow deep neural network structure and PCE, to tackle the limitation of DBT diagnosis, we develop a PCE-based DBT Dual-Net model, which can effectively integrate both low and high-level features for various-sized breast lesions.

2 Methodology

Figure 1 illustrates the proposed DBT lesion classification framework, which consists of two stages: PCE and the application of the Dual-Net model. First, pseudo-color patch augmentation was applied, where the original grayscale patches were transformed in the HSV color space to enhance discriminative features while preserving structural integrity. Second, a dual-branch classification network processed each augmented patch to extract both shallow and deep features. Based on the extracted features, the network predicted patch-wise probabilities for benign and malignant classification, which were aggregated into a final image-level prediction through a majority voting mechanism to ensure robustness against inter-slice variability.

Figure 1

2.1 Pseudo-color enhancement

Inspired by (15), we propose a pseudo-color patch enhancement method to improve the visual visibility of DBT lesion regions, as shown in Figure 1a. We applied a transformation based on the HSV color space and conducted comparative experiments to assess its impact on classification performance. The enhancement operation was to enhance hue, saturation, and brightness value.

After preprocessing, each input sample consists of grayscale patches, denoted as , where . Each single-channel patch undergoes a three-step transformation: conversion to HSV, channel-specific enhancement and back to RGB.

The enhancement procedure modifies HSV channels as follows:

Figure 2 visualizes the differences between the original and pseudo color-enhanced patches. After mapping the grayscale image to color, compared to Figures 2a, b shows clearer irregular edges of the tumor, while Figure 2c exhibits blurred tumor edges but more prominent internal features. For comparison, the Figure 2d shows the image after histogram equalization. To find the optimal HSV adjustment parameter, Figure 3 highlights the impact of HSV-based adjustments on the lesion area and surrounding tissues. After HSV-based adjustment, the image enhances tumor boundary contrast while avoiding the excessive internal contrast seen in the original HSV image, which can lead to loss of subtle texture details. Based on this, the enhancement parameters were empirically set to , , and in the following experiment.

Figure 2

Figure 3

2.2 DBT Dual-Net

2.2.1 Dual-branch feature extraction

The dual-branch included a shallow network branch and a deep network branch . They were combined, and the backbone of their feature extraction parts was retained. The categorization of networks into shallow or deep architectures was based on parameter counts and layer depths, with specific classification criteria presented in Table 1.

The shallow and deep branches are distinguished based on the parameter and the number of layers, summarized in Table 1. Shallow networks ( layers), such as AlexNet and ResNet18, capture fine anatomical details, whereas deep networks (>50 layers), such as ResNet50, ResNeXt50, DenseNet121 and DenseNet201, excel at extracting high-level semantic features.

The deep branch outputs a global feature vector , while the shallow branch outputs a spatial feature map . An adaptive pooling operation is applied to unify spatial dimensions, followed by a flattening operation The vectors and are projected into a shared -dimensional space via respectively. Then, the projected features are used for the similarity loss to encourage complementary representations. The original features are concatenated to form: which is passed to the classifier.

2.2.2 Image-level prediction

Majority voting—For an image consisting of patches , patch-level predictions are aggregated into an image-level label via majority voting, as shown in Equation 5:This aggregation mitigates inter-slice variability and leverages the spatial heterogeneity of lesion appearance in DBT volumes.

Average voting—As a comparison, this approach computes the mean predicted probability across all patches in a case, then applies a threshold of to determine the final case label, as shown in Equation 6:where is the predicted slice-level probability score.

2.2.3 Loss function

To help the two branches to capture representations efficiently, we employed a joint loss that combined the standard binary cross-entropy (BCE) classification loss with a similarity loss between the projected features of the two branches. The total loss was defined as a weighted sum of these components and averaged across all slices in the batch.

For each input slice , the model produced a predicted probability . The BCE loss for slice is computed as shown in Equation 7:where denotes the ground truth label.

In addition, to encourage the dual branches to extract feature representations that are, to some extent, consistent. For this, we introduce a similarity loss between the projected features and of slice . This loss was based on mean squared error, as shown in Equation 8:where and are the outputs of linear projections applied to the features from each branch.

Finally, the overall training objective was formulated as shown in Equation 9:where is the number of slices in the batch and is a balancing coefficient that controls the contribution of the similarity term.

2.3 Dataset

For this study, we utilized the public dataset from the Breast Cancer Screening-Digital Breast Tomosynthesis (BCS-DBT) dataset, publicly available in the Cancer Imaging Archive (39). This dataset was obtained from Duke Health and made available in anonymized DICOM format. It consists of 22,032 breast tomosynthesis scans from 5,060 patients. The spatial resolution is 0.085 mm in the XY direction and 1 mm in the Z direction. The corresponding label files of the dataset contain annotations made by radiologists for each case, including the patient ID, the type of lesion (actionable, benign, and malignant), the imaging view associated with the case, the specific center slice number and position (bounding box) of the lesion. This dataset splits training, validation, and testing sets at the patient level. The distribution of malignant and benign cases, which are the ones used in this study, is shown in Table 2.

Figures 4, 5 present the distribution of lesion width, height, and area in the training set, highlighting the considerable variability in lesion size. The lesions range from approximately 5 mm to 80 mm in width and height, with areas spanning from about 100  to over 3,000 . Such large variations in lesion size may negatively impact the performance of deep learning–based diagnostic models. Figure 6 presents some selected examples for visual comparison. We preprocessed the images as follows:

Figure 4

Figure 5

Figure 6

2.3.1 Lesion ROI extraction

Since DBT data contains images acquired from different views, all images were first uniformly flipped to a left-oriented alignment to standardize the dataset. For lesion region extraction, our processing pipeline implemented multi-axis standardization. In the XY-plane, lesion regions were extracted using annotated bounding box coordinates, with each box’s width and height expanded by 10 pixels in all directions (a total of 20-pixel expansion per dimension) to preserve the surrounding parenchymal context. We analyzed lesion thickness in every slice. We found that the smallest tumor had 10 mm diameter. By assuming lesions are roughly spherical, we selected 10 slices, equivalent to 10 mm, along the Z-axis around the slice that contained the annotation. In this way, we make sure that all considered slices contain tumoral tissue. Besides, the ROI intensities were normalized to the range , and the cropped regions were resized to pixels.

2.3.2 Patch sampling

After lesion extraction, for each DBT lesion case, a minimum 10 mm was initially available across the Z-direction. We employed a center-slice-based symmetrical sampling approach: we extracted different slices at symmetric positions around the lesion center. Figure 7 visualizes the smallest lesion case. As shown, all slices contain tissue with the lesion. To evaluate how different numbers of slice sampling affect diagnostic accuracy, we conducted controlled experiments in which both the training and validation phases used identical slice quantities per case.

Figure 7

2.3.3 Data augmentation

3 Experimental results

3.1 Experimental setting

All models were developed using PyTorch and trained on NVIDIA A100 GPUs. We employed the Adam optimizer. The choice of learning rates was based on empirical selection within the range of to with a 10 times increase. The learning rate of was the best. The training process consisted of 150 epochs for complete model convergence.

3.2 Performance measure

We evaluate classifiers using standard metrics: Accuracy (overall correctness), area under the ROC curve AUC, precision (correctness among predicted positives), sensitivity (fraction of actual positives detected), specificity (fraction of actual negatives correctly identified), and F1 (harmonic mean of precision and recall).

3.3 Results

First, we performed a sensitivity analysis on the different parameter settings of the proposed method. Finally, we compared the proposed method with competing classification methods.

3.3.1 Parameter selection

3.3.1.1 Combinations of different shallow and deep networks

Table 3 shows the performance of the different combinations of shallow and deep networks in Dual-Net. Figure 8 also shows the ROC curves of the different combinations. As shown, except for precision and specificity, the combination of AlexNet and ResNet50 gave the best results. For this reason, this combination was used in the comparison with other methods. It is interesting to see that pure CNN-based combinations outperform hybrid methods that include ViTs.

Figure 8

3.3.1.2 Weight factor at similarity loss

Concerning the hyperparameter of the similarity loss, we tested the performance of Dual-Net with different values. The was designed to constrain similarity between two features extracted from each branch. The result is shown in Table 4, which shows that has the best performance for AUC, which is usually the preferred metric in clinical settings. Disregarding the similarity constraint also gives competitive results.

3.3.1.3 Effect of pseudo-color enhancement

To validate the effectiveness of PCE, we compare the classification performance without image enhancement, with PCE, and with histogram equalization. As shown in Table 5, incorporating PCE yields the best performance, while histogram equalization also improves performance but to a lesser extent. Specifically, accuracy increased from 65.94% to 75.36%, and AUC rose from 69.45 to 77.80, indicating better overall discrimination. Precision, specificity, and F1-score also improved substantially, reflecting more balanced performance across classes. We compare to the without and the best lambda weight factor with the similarity loss.

3.3.1.4 Different numbers of adjacent slices

We also compared the classification performance at the image level with different numbers of adjacent slices as input. We choose and ultimately form a total of slices to predict the image-level results, which serve as the input. The results obtained from voting with different numbers of patch inputs, using the previously best Dual-Net model for this evaluation shown in Table 6. The results show that acquired the optimal classification performance (i.e., the center slice plus the adjacent slices), but incorporating excessive slice information does not necessarily improve tumor characterization performance.

3.3.2 Comparison with previous methods

To demonstrate superiority of the proposed method, we compared the proposed method with different fine-tuned deep learning classifiers, including ResNet (40), AlexNet (41), ResNeXt50 (42), MobileNetV2 (43), DenseNet (44), ViT (45) with pretrained weights. These classifiers were selected as they cover architectures with different depths and complexities. Moreover, a feature pyramid network (FPN) (46) with ResNet50 backbone was also employed for performance comparison. This architecture integrates feature maps from four different convolution layers, which are subsequently concatenated before making final predictions. We performed a five-fold cross-validation in the experiments.

3.3.2.1 Classification performance at slice level

Table 7 compares the performance of different network architectures for DBT lesion classification, and Figure 9 presents the ROC curves of the corresponding models. The results do not clearly that either shallow or deep networks are better for DBT classification However, our proposed method (Dual-Net) results have the best performance.

Figure 9

Figure 10

3.3.2.2 Classification performance at image level

To obtain image-level predictions from DBT volumes, a voting-based aggregation strategy was applied. Table 8 and Figure 10 report the corresponding image-level classification performance. Each image contains patches that share a single ground-truth label. Two aggregation methods were evaluated: majority voting () and average voting () denoted in Table 8. For AUC and other threshold-independent metrics, the image-level predicted probabilities are from the averaged selected patches. As shown in Table 8, except for precision, Dual-Net was the best method. The majority voting strategy focuses on the consensus among patch-level predictions, whereas average voting incorporates the confidence of each patch prediction into the aggregation process.

3.3.2.3 Statistical analysis

To validate the performance improvements, the final version of Dual-Net was used to compare with the best-performing slice-level and image-level models in terms of AUC. A paired per-case analysis was conducted, and statistical significance was assessed via McNemar tests with 2 2 contingency tables. Table 9 shows that Dual-Net significantly outperforms the best baselines ().

All comparisons yielded statistically significant differences (), confirming the robustness of our final Dual-Net architecture and its superiority across diverse configurations and preprocessing approaches.

3.3.3 Visualization results

3.3.3.1 t-SNE Visualization

To investigate why Dual-Net outperforms the single-branch network in classification, t-distributed Stochastic Neighbor Embedding (t-SNE) (47) is used. We separately extracted the fused features from the dual-branch network and the features from the single-branch network. After dimensionality reduction, the benign and malignant features for were used for comparison. As shown in Figure 11, it will be challenging for classifiers to disentangle the two classes based on the features generated by single-branch networks. In turn, even a very simple classifier could perform well on features extracted with Dual-Net since it exhibits clearer inter-class separation and intra-class clustering.

Figure 11

3.3.3.2 Grad-CAM visualization

In addition to quantitative comparisons, we conducted explainability analysis to understand model behaviors. We visualized their attention patterns using Grad-CAM (48) and overlaid these heatmaps on the original images.

For Dual-Net, we independently extract feature maps from the final convolutional layer of each branch. Following the computation of gradients on these feature maps, a direct summation strategy to fuse the weighted feature maps from both branches, ultimately projecting the results onto the original image through chromatic mapping. As shown in Figure 12, a representative case illustrates how our model integrates complementary attention patterns from both branches. In this example, we can see that, unlike DenseNet121, Dual-Net focuses on anatomically plausible areas. Thus, the dual-branch architecture attended to more clinically relevant regions.

Figure 12

4 Discussion

In this work, we proposed a PCE-based Dual-Net to classify lesions in DBT. The results show the effectiveness of PCE pre-processing in DBT diagnosis. This is in line with several studies that have shown that PCE can improve the medical imaging diagnosis performance (35, 36). We hypothesize that PCE alters the distribution of intensity, which enables the model to make more accurate diagnoses. Psychophysical studies have shown that humans can distinguish between approximately 1–10 million colors, but only around 1,000 gray scales (49). Thus, humans can benefit from using PCE. In clinical practice, pseudo-color can address the shortcomings of grayscale displays and has been widely applied in various clinical digital image analyses (50). For example, PCE can markedly enhance the morphological information of maxillary tumors in oral images and make hidden, millet-sized lesions in thyroid spiral CT images clearly visible. Other applications include multimodal overlays, such as overlaying pseudo-colored PET on CT anatomy to conveniently display functional–structural relationships. While PCE is beneficial for visual interpretation performed by humans, the advantages of using PCE as a pre-processing step for deep learning models are less clear. Indeed, any potential mapping to pseudo-colors would become embedded within the neural network, and, in itself, PCE does not provide additional information that could be useful for the model. Still, our results show that PCE is effective, aligning with previous studies. Even simpler image enhancement methods, such as histogram equalization, yielded improved results. We hypothesize that PCE is effective in cases with limited training data, which is typically the case in medical imaging applications. For sufficiently large datasets, the network can learn richer representations, where pre-processing steps, such as PCE, are expected to have a lesser impact.

4.1 Limitations

There are several critical aspects for further investigation. First, the selection of an appropriate backbone network for Dual-Net is crucial for performance. Through comparative experiments, we found that the combination of AlexNet and ResNet achieved the best performance. The result is consistent not only with our experimental observations but also with previous studies, which highlight the importance of low-level features from shallow networks in computer vision tasks (51, 52). In terms of model architecture, we fused two models to capture both low-level and high-level representations, leveraging the complementary strengths of shallow and deep networks. The combinations of CNNs-ViTs and ViT models are also used for comparison. However, methods that included ViTs did not outperform CNN-only models. We attribute it to the limited size of our dataset. Compared to ViT, CNNs have local inductive biases that might help them converge faster in datasets with a limited amount of data (53–55). While the AlexNet–ResNet model performed well on the datasets, its generalization remains constrained by dataset limitations. Future work will focus on expanding the dataset to validate robustness, which may also unlock the potential of more complex architectures, such as ResNet–ViT (54).

Second, at the image level prediction, since each lesion is only annotated with a single slice, we assume each lesion region is spherical in shape. This assumption is common in DBT imaging evaluations (56, 57) or physical phantom simulations (58). However, several studies report that only about 18% of tumors are approximately spherical (53, 60), while many present as discoid or irregular shapes. Thus, our assumption of spherical lesion shape is not fully accurate. However, its impact on the results is expected to be minimal. This is because we chose the smallest lesion in the dataset–10 mm in our case–to estimate the number of slices that are considered for processing in all cases, which corresponds to 10 slices. The probability of having slices without lesion tissue largely decreases with the size of the lesion. Furthermore, in DBT, reconstruction artifacts typically make lesion regions appear larger than their actual size. We manually assessed 10 slices close to the smallest lesions and found lesion tissue in all of them. Figure 7, which visualizes the slices surrounding the smallest lesion, shows that all slices contain lesion tissue. We uniformly set the maximum lesion thickness to 10 slices, determined by the smallest lesion thickness in the dataset. Although this choice does not perfectly reflect the true distribution of lesion sizes, it allowed us to maximize data utilization in a dataset where lesion samples are relatively scarce. In future studies with larger datasets, more flexible sampling strategies that better capture thickness variability will be considered.

Last but not least, we acknowledge some methods from the literature not included for comparison. This limitation stems from two practical considerations: (1) many referenced methods were tested on private datasets, making fair comparisons infeasible; (2) incomplete implementation details in existing works hinder faithful reproduction of their results. At the moment, public DBT data is scarce. In the future, we will validate our method with additional datasets.

5 Conclusion

In summary, we proposed a Dual-Net model for DBT lesion classification, integrating a dual-branch network feature leveraging complementary representations from two distinct networks. We systematically evaluated the performance of different deep learning architectures for the classification of benign and malignant lesions in DBT by using the public BCS-DBT dataset. Under identical experimental settings, our approach demonstrated superior classification performance compared to baseline models.

References

• 1.

  Sung H Ferlay J Siegel RL Laversanne M Soerjomataram I Jemal A 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. 10.3322/caac.21660

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 2.

  El-Banby GM Salem NS Tafweek EA Abd El-Azziz EN . Automated abnormalities detection in mammography using deep learning. Complex Intell Syst. (2024) 10(5):7279–95. 10.1007/s40747-024-01532-x

  • CrossRef
  • Google Scholar

• 3.

  Niklason LT Christian BT Niklason LE Kopans DB Castleberry DE Opsahl-Ong BH et al Digital tomosynthesis in breast imaging. Radiology. (1997) 205(2):399–406. 10.1148/radiology.205.2.9356620

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 4.

  Yerushalmi R Woods R Ravdin PM Hayes MM Gelmon KA . Ki67 in breast cancer: prognostic and predictive potential. Lancet Oncol. (2010) 11(2):174–83. 10.1016/s1470-2045(09)70262-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 5.

  Chen X Zhang Y Zhou J Wang X Liu X Nie K , et al. Diagnosis of architectural distortion on digital breast tomosynthesis using radiomics and deep learning. Front Oncol. (2022) 12:991892. 10.3389/fonc.2022.991892

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 6.

  Oladimeji OO Ayaz H McLoughlin I Unnikrishnan S . Mutual information-based radiomic feature selection with SHAP explainability for breast cancer diagnosis. Results Eng. (2024) 24:103071. 10.1016/j.rineng.2024.103071

  • CrossRef
  • Google Scholar

• 7.

  Tagliafico AS Valdora F Mariscotti G Durando M Nori J La Forgia D , et al. An exploratory radiomics analysis on digital breast tomosynthesis in women with mammographically negative dense breasts. Breast. (2018) 40:92–6. 10.1016/j.breast.2018.04.016

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 8.

  Sakai A Onishi Y Matsui M Adachi H Teramoto A Saito K , et al. A method for the automated classification of benign and malignant masses on digital breast tomosynthesis images using machine learning and radiomic features. Radiol Phys Technol. (2020) 13:27–36. 10.1007/s12194-019-00543-5

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 9.

  Jiang T Jiang W Chang S Wang H Niu S Yue Z , et al. Intratumoral analysis of digital breast tomosynthesis for predicting the Ki-67 level in breast cancer: a multi-center radiomics study. Med Phys. (2022) 49:219–30. 10.1002/mp.15392

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 10.

  Son J Lee SE Kim E-K Kim S . Prediction of breast cancer molecular subtypes using radiomics signatures of synthetic mammography from digital breast tomosynthesis. Sci Rep. (2020) 10:21566. 10.1038/s41598-020-78681-9

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 11.

  Tagliafico AS Bignotti B Rossi F Matos J Calabrese M Valdora F , et al. Breast cancer Ki-67 expression prediction by digital breast tomosynthesis radiomics features. Eur Radiol Exp. (2019) 3:1–6. 10.1186/s41747-019-0117-2

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 12.

  Yang Z Fan T Smedby Ö Moreno R . Lesion localization in digital breast tomosynthesis with deformable transformers by using 2.5D information. In: Medical Imaging 2024: Computer-Aided Diagnosis. SPIE (2024). Vol. 12927, p. 84–9. 10.1117/12.3005496

  • CrossRef
  • Google Scholar

• 13.

  Alashban Y . Breast cancer detection and classification with digital breast tomosynthesis: a two-stage deep learning approach. Diagn Interv Radiol. (2025) 31:206. 10.4274/dir.2024.242923

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 14.

  Du Y Hooley RJ Lewin J Dvornek NC . SIFT-DBT: self-supervised initialization and fine-tuning for imbalanced digital breast tomosynthesis image classification. In: 2024 IEEE International Symposium on Biomedical Imaging (ISBI). (2024). p. 1–5. 10.1109/ISBI56570.2024.10635723

  • CrossRef
  • Google Scholar

• 15.

  El-Shazli AMA Youssef SM Soliman AH . Intelligent computer-aided model for efficient diagnosis of digital breast tomosynthesis 3D imaging using deep learning. Appl Sci. (2022) 12:5736. 10.3390/app12115736

  • CrossRef
  • Google Scholar

• 16.

  Kassis I Lederman D Ben-Arie G Giladi Rosenthal M Shelef I Zigel Y . Detection of breast cancer in digital breast tomosynthesis with vision transformers. Sci Rep. (2024) 14:22149–12. 10.1038/s41598-024-72707-2

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 17.

  Lee W Lee H Lee H Park EK Nam H Kooi T . Transformer-based deep neural network for breast cancer classification on digital breast tomosynthesis images. Radiol Artif Intell. (2023) 5:e220159. 10.1148/ryai.220159

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 18.

  Li X Qin G He Q Sun L Zeng H He Z , et al. Digital breast tomosynthesis versus digital mammography: integration of image modalities enhances deep learning-based breast mass classification. Eur Radiol. (2020) 30:778–88. 10.1007/s00330-019-06457-5

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 19.

  Ricciardi R Mettivier G Staffa M Sarno A Acampora G Minelli S , et al. A deep learning classifier for digital breast tomosynthesis. Phys Med. (2021) 83:184–93. 10.1016/j.ejmp.2021.03.021

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 20.

  Xiao B Sun H Meng Y Peng Y Yang X Chen S , et al. Classification of microcalcification clusters in digital breast tomosynthesis using ensemble convolutional neural network. Biomed Eng Online. (2021) 20:71. 10.1186/s12938-021-00908-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 21.

  Zhang Y Wang X Blanton H Liang G Xing X Jacobs N . 2D convolutional neural networks for 3D digital breast tomosynthesis classification. In: 2019 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). (2019). p. 1013–7. 10.1109/BIBM47256.2019.8983097

  • CrossRef
  • Google Scholar

• 22.

  Ren H Ge D-F Yang Z-C Cheng Z-T Zhao S-X Zhang B . The value of predicting breast cancer with a DBT 2.5D deep learning model. Discov Oncol. (2025) 16:1–14. 10.1007/s12672-024-01698-3

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 23.

  Mendel K Li H Sheth D Giger M . Transfer learning from convolutional neural networks for computer-aided diagnosis: a comparison of digital breast tomosynthesis and full-field digital mammography. Acad Radiol. (2019) 26:735–43. 10.1016/j.acra.2018.06.019

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 24.

  Samala RK Chan H-P Hadjiiski L Helvie MA Richter CD Cha KH . Breast cancer diagnosis in digital breast tomosynthesis: effects of training sample size on multi-stage transfer learning using deep neural nets. IEEE Trans Med Imaging. (2018) 38:686–96. 10.1109/TMI.2018.2870343

  • CrossRef
  • Google Scholar

• 25.

  Samala RK Chan H-P Hadjiiski LM Helvie MA Richter C Cha K . Evolutionary pruning of transfer learned deep convolutional neural network for breast cancer diagnosis in digital breast tomosynthesis. Phys Med Biol. (2018) 63:095005. 10.1088/1361-6560/aabb5b

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 26.

  Wang L Zheng C Chen W He Q Li X Zhang S , et al. Multi-path synergic fusion deep neural network framework for breast mass classification using digital breast tomosynthesis. Phys Med Biol. (2020) 65:235045. 10.1088/1361-6560/abaeb7

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 27.

  Yang Z Fan T Smedby Ö Moreno R . 3D breast ultrasound image classification using 2.5D deep learning. In: 17th International Workshop on Breast Imaging (IWBI 2024). SPIE (2024). Vol. 13174. p. 443–9. 10.1117/12.3025534

  • CrossRef
  • Google Scholar

• 28.

  Welch HG Prorok PC O’Malley AJ Kramer BS . Breast-cancer tumor size, overdiagnosis, and mammography screening effectiveness. N Engl J Med. (2016) 375(15):1438–47. 10.1056/NEJMoa1600249

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 29.

  Thakur GK Thakur A Kulkarni S Khan N Khan S . Deep learning approaches for medical image analysis and diagnosis. Cureus. (2024) 16:e59507. 10.7759/cureus.59507

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 30.

  Kaya Y Hong S Dumitras T . Shallow-deep networks: understanding and mitigating network overthinking. In: International Conference on Machine Learning. PMLR (2019). Vol. 97. p. 3301–10. Available online at: https://proceedings.mlr.press/v97/kaya19a.html

  • Google Scholar

• 31.

  Robles Herrera S Ceberio M Kreinovich V . When is deep learning better and when is shallow learning better: qualitative analysis. Int J Parallel Emergent Distrib Syst. (2022) 37:589–95. 10.1080/17445760.2022.2070748

  • CrossRef
  • Google Scholar

• 32.

  Jian M Wu R Chen H Fu L Yang C . Dual-branch-UNet: a dual-branch convolutional neural network for medical image segmentation. Comput Model Eng Sci. (2023) 137:705–16. 10.32604/cmes.2023.027425

  • CrossRef
  • Google Scholar

• 33.

  Gao F Wu T Li J Zheng B Ruan L Shang D , et al. SD-CNN: a shallow-deep CNN for improved breast cancer diagnosis. Comput Med Imaging Graph. (2018) 70:53–62. 10.1016/j.compmedimag.2018.09.004

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 34.

  Zhao G Feng Q Chen C Zhou Z Yu Y . Diagnose like a radiologist: hybrid neuro-probabilistic reasoning for attribute-based medical image diagnosis. IEEE Trans Pattern Anal Mach Intell. (2021) 44:7400–16. 10.1109/TPAMI.2021.3130759

  • CrossRef
  • Google Scholar

• 35.

  Huang X Wang Q Chen J Chen L Chen Z . Effective hybrid attention network based on pseudo-color enhancement in ultrasound image segmentation. Image Vis Comput. (2023) 137:104742. 10.1016/j.imavis.2023.104742

  • CrossRef
  • Google Scholar

• 36.

  Zhang Z Jiang H Liu J Shi T . Improving the fidelity of CT image colorization based on pseudo-intensity model and tumor metabolism enhancement. Comput Biol Med. (2021) 138:104885. 10.1016/j.compbiomed.2021.104885

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 37.

  Hussein AA Valizadeh M Amirani MC Mirbolouk S . Breast lesion classification via colorized mammograms and transfer learning in a novel CAD framework. Sci Rep. (2025) 15:25071. 10.1038/s41598-025-10896-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 38.

  Jones MA Zhang K Faiz R Islam W Jo J Zheng B , et al. Utilizing pseudo color image to improve the performance of deep transfer learning–based computer-aided diagnosis schemes in breast mass classification. J Imaging Inform Med. (2025) 38:1871–80. 10.1007/s10278-024-01237-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 39.

  Buda M Saha A Walsh R Ghate S Li N Swiecicki A et al Breast cancer screening - digital breast tomosynthesis (BCS-DBT). Cancer Imaging Arch. (2020). Version 5. 10.7937/E4WT-CD02

  • CrossRef
  • Google Scholar

• 40.

  He K Zhang X Ren S Sun J . Deep residual learning for image recognition. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). (2016). p. 770–8. 10.1109/CVPR.2016.90

  • CrossRef
  • Google Scholar

• 41.

  Krizhevsky A Sutskever I Hinton GE . ImageNet classification with deep convolutional neural networks. In: Advances in Neural Information Processing Systems (NIPS 2012). (2012) Vol. 25. p. 1–9. Available online at:https://papers.nips.cc/paper_files/paper/2012/hash/c399862d3b9d6b76c8436e924a68c45b-Abstract.html

  • Google Scholar

• 42.

  Xie S Girshick R Dollár P Tu Z He K . Aggregated residual transformations for deep neural networks. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). (2017). p. 1492–500. 10.1109/CVPR.2017.634

  • CrossRef
  • Google Scholar

• 43.

  Sandler M Howard A Zhu M Zhmoginov A Chen L-C . MobileNetV2: inverted residuals and linear bottlenecks. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). (2018). p. 4510–20. 10.1109/CVPR.2018.00474

  • CrossRef
  • Google Scholar

• 44.

  Huang G Liu Z Van Der Maaten L Weinberger KQ . Densely connected convolutional networks. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). (2017). p. 4700–8. 10.1109/CVPR.2017.243

  • CrossRef
  • Google Scholar

• 45.

  Dosovitskiy A Beyer L Kolesnikov A Weissenborn D Zhai X Unterthiner T , et al. An image is worth 16 16 words: transformers for image recognition at scale. In: International Conference on Learning Representations (ICLR). (2021). Available online at: https://openreview.net/forum?id=YicbFdNTTy

  • Google Scholar

• 46.

  Lin T-Y Dollár P Girshick R He K Hariharan B Belongie S . Feature pyramid networks for object detection. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). (2017). p. 2117–25. 10.1109/CVPR.2017.106

  • CrossRef
  • Google Scholar

• 47.

  Maaten LVD Hinton G . Visualizing data using t-SNE. J Mach Learn Res. (2008) 9:2579–605. Available online at: https://www.jmlr.org/papers/v9/vandermaaten08a.html

  • Google Scholar

• 48.

  Selvaraju RR Cogswell M Das A Vedantam R Parikh D Batra D . Grad-CAM: visual explanations from deep networks via gradient-based localization. In: IEEE International Conference on Computer Vision (ICCV). (2017). p. 618–26. 10.1109/ICCV.2017.74

  • CrossRef
  • Google Scholar

• 49.

  Pointer MR Attridge GG . The number of discernible colours. Color Res Appl. (1998) 23(1):52–4. 10.1002/(SICI)1520-6378(199802)23:1%3C52::AID-COL8%3E3.0.CO;2-2

  • CrossRef
  • Google Scholar

• 50.

  Lan C Lan P Cao Y . Pseduo-colour technique in the medical images. In: Proceedings of the 4th China-Japan-Korea Joint Symposium on Medical Informatics. Beijing. (2002). p. 180–2.

  • Google Scholar

• 51.

  Cerra D Pato M Carmona E Azimi SM Tian J Bahmanyar R , et al. Combining deep and shallow neural networks with ad hoc detectors for the classification of complex multi-modal urban scenes. In: 2018 IEEE International Geoscience and Remote Sensing Symposium (IGARSS). (2018). p. 3856–9. 10.1109/IGARSS.2018.8517699

  • CrossRef
  • Google Scholar

• 52.

  Sun X Lv M . Facial expression recognition based on a hybrid model combining deep and shallow features. Cognit Comput. (2019) 11:587–97. 10.1007/s12559-019-09654-y

  • CrossRef
  • Google Scholar

• 53.

  Bendazzoli S Tzortzakakis A Abrahamsson A Engelbrekt Wahlin B Smedby Ö Holstensson M et al Anatomy-Aware Lymphoma Lesion Detection in Whole-Body. arXiv preprint 2025). 10.48550/arXiv.2511.07047

  • CrossRef
  • Google Scholar

• 54.

  Khan A Rauf Z Sohail A Khan AR Asif H Asif A , et al. A survey of the vision transformers and their CNN-transformer based variants. Artif Intell Rev. (2023) 56:2917–70. 10.1007/s10462-023-10595-0

  • CrossRef
  • Google Scholar

• 55.

  Maurício J Domingues I Bernardino J . Comparing vision transformers and convolutional neural networks for image classification: a literature review. Appl Sci. (2023) 13:5521. 10.3390/app13095521

  • CrossRef
  • Google Scholar

• 56.

  Bakic P Barufaldi B Higginbotham D Weinstein SP Avanaki AN Espig KS , et al. Virtual clinical trial of lesion detection in digital mammography and digital breast tomosynthesis. In: SPIE Medical Imaging: Physics of Medical Imaging, 1057306. (2018) Vol. 10573. 10.1117/12.2294934

  • CrossRef
  • Google Scholar

• 57.

  Reiser I Bian J Nishikawa R Sidky E Pan X . Comparison of reconstruction algorithms for digital breast tomosynthesis. In: International Meeting on Fully Three-Dimensional Image Reconstruction in Radiology and Nuclear Medicine (Fully 3D). (2007). 10.48550/arXiv.0908.2610

  • CrossRef
  • Google Scholar

• 58.

  Zimmermann BB Deng B Singh B Martino M Selb J Fang Q , et al. Multimodal breast cancer imaging using coregistered dynamic diffuse optical tomography and digital breast tomosynthesis. J Biomed Opt. (2017) 22:046008. 10.1117/1.JBO.22.4.046008

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 59.

  Byrd BK Krishnaswamy V Gui J Rooney T Zuurbier R Rosenkranz K , et al. The shape of breast cancer. Breast Cancer Res Treat. (2020) 183:403–10. 10.1007/s10549-020-05780-6

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 60.

  Wapnir IL Wartenberg DE Greco RS . Three dimensional staging of breast cancer. Breast Cancer Res Treat. (1996) 41:15–9. 10.1007/BF01807032

  • Pubmed Abstract
  • CrossRef
  • Google Scholar