An Adaptive Cycle-GAN-Based Augmented LIME-Enabled Multi-Stage Transfer Learning Model for Improving Breast Tumor Detection Using Ultrasound Images

Abstract

Breast cancer is recognized as an aggressive cancer with the highest rate of mortality. Ultrasound imaging is a non-invasive and cost-effective strategy which is most frequently utilized in clinical methods. Especially, in ultrasound scan, breast tumors may appear in blurred and unclear boundaries. Thus, there is a necessity to improve the quality of breast ultrasound images. In this work, we introduce a cycle generative adversarial network (GAN) for translating noisy breast ultrasound images to denoised images. Furthermore, translating denoised images to reconstructed images helps in preserving breast tumor boundaries for better efficacy. To accurately identify the augmented breast tumor images, we consider an ensemble model of pre-trained transfer learning models such as Inception-v3, Densenet121, and XceptionLike. Furthermore, we present an automated boundary extraction using Local Interpretable Model-agnostic Explanations (LIME), providing interpretability for boundary extraction in breast lesions from ultrasound images. Through experimentation, we have achieved 93% of accuracy for the proposed model, and LIME provides better interpretability for each pre-trained model. Furthermore, the proposed model outperforms Vison Transformer (ViT) models.

1. Introduction

Breast cancer has emerged as the most common malignancy in the globe, with more than 2 million new cases since 2020 [1]. The most accurate way to effectively handle breast cancer is to diagnose the cancer at an early stage. Magnetic resonance imaging (MRI), Breast ultrasound (BUS) images, and computed tomography (CT) are non-invasive strategies for early breast tumor identification. However, BUS is effective for soft tissues and less expensive than MRI and CT. In addition, BUS is an effective approach to initially diagnose benign and malignant breast tumors. Moreover, BUS does not require ionizing radiations and is more effective in diagnosing dense breast tissues, whereas mammograms are very sensitivity and less effective in dense tissues [2]. In the medical field, automated diagnostic systems have emerged as vital for breast cancer identification. Moreover, automated diagnostic systems may help physicians in diagnosing cancer patients and reduce the mortality rate through remote medical diagnosis [3]. For BUS images, image pre-processing and segmentation techniques are important for accurate classification. Most of the existing studies rely on texture changes and image segmentation techniques which may not retrieve essential features and suffer from overfitting [4,5]. However, BUS images may have blurred, low contrast, speckle noise and unclear boundaries, which makes early and accurate diagnosis challenging [6]. Thus, there is a necessity to improve the quality of breast ultrasound images.

Even though medical datasets are publicly available, the majority of datasets exhibit an imbalanced class distribution and are limited in size. However, acquiring health-related data is a time-consuming and expensive process that relies on collaborating with radiologists and clinical researchers [7]. Recently, for BUS images, generative learning has attained superior performance in comparison to supervised learning [8]. A new era of generative learning with the rise of generative adversarial networks (GANs) has gained attention, generating high quality synthetic BUS images. Some existing studies have adapted the GAN for medical image segmentation, generating realistic texture translation and medical image denoising [9,10]. In this paper, we have considered an adaptive cycle-GAN to eliminate unwanted distortion from BUS images. Moreover, the proposed adaptive cycle-GAN generates high-quality BUS images from unwanted distortion of low-quality BUS images in order to improve diagnostic accuracy. In addition, the performance of the proposed adaptive cycle-GAN is evaluated based on the loss function, which relies on cycle consistency loss, adversarial loss, and the residual loss of encoder and decoder networks for image denoising.

For deep learning applications, the datasets with a sufficient number of training samples and credible ground truth are essential factors for an accurate diagnosis [11]. For a medical dataset, acquiring these factors is a major difficult task. Thus, transfer learning helps in utilizing the knowledge that has been trained on a particular model in order to handle a similar task [12]. In addition, transfer learning also helps to improve the performance, especially when the datasets are limited in size. Recently, some existing studies have depicted that pre-trained convolutional neural network (CNN) models have attained higher accuracy, especially with limited datasets [13,14]. However, during transfer learning, identifying the layers of pre-trained model which require fine-tuning is a challenging task. This problem can be addressed with the consideration of an ensemble of transfer learning through an adaptive fine-tuning techniques (like freezing early layers) that can reduce the over-fitting and computational cost. In this work, we consider an ensemble of multi-stage transfer learning models such as DenseNet 121, Inception-V3, and XceptionLike for classifying breast tumor images. The predictions of the proposed ensemble-based multi-stage transfer learning model are explained using explainable artificial intelligence techniques such as LIME. For breast tumor diagnosis, LIME provides an individual predictions on the local interpretable representation of BUS images. Thus, with the consideration of LIME, the efficacy of the proposed model improves trustworthiness in diagnosing breast cancer patients. In this work, the major contributions are summarized as follows:

• We have considered an adaptive cycle-GAN to eliminate unwanted distortion from BUS images.
• We consider an ensemble of multi-stage transfer learning models such as DenseNet 121, Inception-V3, and XceptionLike for classifying breast tumor images.
• Furthermore, we present an automated boundary extraction using Local Interpretable Model-agnostic Explanations (LIME) which highlights the boundary extraction of breast lesions from BUS images and provides the interpretable explanations on the predictions of the proposed multi-stage transfer learning model.
• During experimentation, the proposed model’s efficacy has been assessed and compared with other existing approaches.

2. Related Work

In this section, GAN-based synthetic medical image augmentation, identification, and classification of breast tumor using deep techniques have been discussed.

Most of the current existing studies have employed generative adversarial networks (GANs) for medical image augmentation. Recently, some studies have adapted GAN-based image-to-image translation, cross-domain image translation and segmentation to image translation. Xue et al. [15] have considered two adversarial networks, namely, segmentor and critic for medical image segmentation. Dai et al. [16] have adapted the GAN to generate realistic segmentation that resembles the ground truth. In [17], the authors have employed semi-supervised the GAN for breast ultrasound image segmentation. Schlegl et al. [18] have applied the GAN on the retinal region for the identification of benign tissues. Furthermore, the authors have validated the GAN for the identification of anomalous data in retinal images. In [19], the authors have trained their model using the GAN for data augmentation to improve breast ultrasound images. However, the model cannot achieve better performance with unknown noisy data in real images. Zhong et al. [20] have adapted the GAN for image denoising using convolutional neural networks (CNNs). Moreover, neural networks often suffer from data drift and vanishing or exploding gradient problems. To avoid data drift, the residual loss of an autoencoder within a generator network of the GAN helps in extracting the features from the input BUS images and improves the training stability of the GAN. Han et al. [21] have adapted an unsupervised GAN model to detect medical anomalies, either from reconstruction loss or the latent feature space. The recent advances in variational autoencoders with different variants of GANs help in improving the efficacy of diagnosing breast cancer patients.

Physicians require a significant amount of time and clinical effort for the precise identification and diagnosis of breast cancer from medical imaging modalities. Moreover, the usage of current diagnostic techniques and software helps in reducing the costs in diagnosing breast cancer. Several existing studies have been proposed for breast cancer detection and diagnosis. In breast tumor diagnosis, deep learning-based frameworks have shown remarkable success in the last few decade. Feng et al. [6] have presented a boundary enhancement for BUS image segmentation using a transformer network that highlights breast lesion region boundaries. In [22], a boundary-rendering network is presented using graph convolutional networks for extracting breast lesions from ultrasound images. In [8], an attentional GAN model is presented for breast tumor identification using BUS images. However, for the multi-classification problem, there exists some challenging issues related to clinical similarities, as the inconsistent staining of biological samples and at different resolutions BUS images from the similar group may exhibit major variations for the interpretation and analysis of medical images. Zheng et al. [23] have presented a deep learning-assisted adaboost algorithm for the early diagnosis and identification of breast cancer. Yala et al. [24] have adapted a reinforcement learning model with a Markov decision process for the early identification of breast cancer. Furthermore, the authors have defined the reward value as assessing the imaging cost and time step between recommended and actual screening dates for early detection. However, deep learning models may not be well tuned due to the limited availability of labeled and annotated medical data. Thus, transfer learning can be applied to improve the performance of deep learning models.

In recent years, transfer learning has been utilized for medical image analysis as computer vision applications have shown to be more beneficial for precise classification. Ayana et al. [25] have adapted transfer learning by considering pre-trained CNN models for the identification of breast cancer from ultrasound images. The authors have utilized two different types of transfer learning models, namely cross-domain and cross-modal models, for breast cancer diagnosis. Wang et al. [26] have developed an automated computer-aided detection system by considering CNNs with a deep transfer learning model, trained on BUS images. The authors have claimed that CNNs with transfer learning have achieved a sensitivity of 89%. In [27], the authors have utilized transfer learning with GoogLeNet and AlexNet for improving mammographic breast lesion identification. In [28], an automated breast tumor identification system is developed by transferring Inceptionv3, ResNet50, and VGG-16. However, in the past decades, studies have yet to explore deep transfer learning in breast imaging. The main objective of this work is to improve the diagnosis of breast cancer with an ensemble of deep transfer learning models using breast imaging modalities.

3. LIME-Assisted Ensemble-Based Multi-Stage Deep Transfer Learning Model

3.1. Cycle GAN

The early and precise identification and diagnosis of cancer from BUS images is a major challenge due to the presence of random noise and blurred images. Hence, BUS image denoising is an essential pre-processing step that helps to improve the quality of the image. Recently, deep learning-based image denoising techniques have been presented for the removal of random noise from BUS imaging modalities. Most of the existing deep learning-based image denoising techniques rely on supervised learning, which requires realistic noisy images with the corresponding ground-truth images. Thus, it may be a difficult task to acquire real-time BUS images. In this work, we consider an unsupervised denoising technique for BUS imaging using a cycle-generative adversarial network (Cycle-GAN). Moreover, the cycle-GAN relies on cycle consistency and the adversarial loss function in order to avoid paired training samples.

As shown in Figure 1, the cycle-GAN mainly employs two generators and two discriminators for distinguishing among real and generated BUS images. The denoising generator network $G_P$ is trained to map from a noisy original medical image P to a distortion-free medical image Q. Moreover, the noisy generator network $F_Q$ is trained by mapping from distortion-free medical image Q to the reconstructed medical image P. As depicted in Figure 2, the generator network contains encoder, transformer, and decoder networks. In generator networks, convolution and deconvolution layers are constructed to represent encoder and decoder networks, respectively. Furthermore, transformers are build using residual blocks that leads to faster convergence by updating their weights. Moreover, each convolution layer is represented by the number of filters and size of filters with ReLU activation and batch normalization. In the deconvolution layer, reconstructed images are generated based on the extracted feature set.

The structure of the discriminator networks is depicted in Figure 3, where BN represents batch normalization. The two discriminator networks, namely $Di{s}_P$ and $Di{s}_Q$, help in distinguishing whether the image contains any distorted data in the medical imaging modalities. The denoised discriminator network $Di{s}_P$ helps to differentiate images from the distortion-free domain, generated by the denoising generator network $G_P$. In addition, the noisy discriminator network $Di{s}_Q$ helps to differentiate images from the reconstructed domain, generated by the noisy generator network $F_Q$. Thus, adversarial loss assures that the generator network helps in generating more genuine BUS images. The generator network $G_P$ maps noisy BUS image X to G(q) (where $q\sim p{r}_D$ represents data distribution) in a distortion-free BUS image. For a given generator network $G\left(q\right)$, the discriminator network $Di{s}_P$ distinguishes whether the image represents the original data in noisy image sample P. The Wasserstein adversarial loss function with two generator networks, $G_P$ and $F_Q$, with gradient penalty is computed as

$$\begin{array}{c}\hfill Los{s}_{WGAN}(Di{s}_P,Di{s}_Q)={\mathbb{E}}_p\left[Di{s}_Q\left(G_P\left(p\right)\right)\right]-{\mathbb{E}}_q\left[Di{s}_Q\left(q\right)\right]+{\mathbb{E}}_q\left[Di{s}_P\left(F_Q\left(q\right)\right)\right]-{\mathbb{E}}_p\left[Di{s}_P\left(p\right)\right]\hfill \\ \hfill +\lambda {\mathbb{E}}_{\widehat{p}}{\left(\parallel {\nabla }_{\widehat{p}}Di{s}_P\left(\widehat{p}\right){\parallel }_2-1\right)}^2+\lambda {\mathbb{E}}_{\widehat{q}}{\left(\parallel {\nabla }_{\widehat{q}}Di{s}_Q\left(\widehat{q}\right){\parallel }_2-1\right)}^2\hfill \end{array}$$

(1)

where the term $\mathbb{E}$ represents the expectation value under probabilistic distribution. The first and second terms in Equation (1) represent the adversarial Wasserstein loss function for $G_P$ and $F_Q$, respectively. Furthermore, the third and fourth terms in Equation (1) represent the gradient penalty for improving stability during model training. However, training two generator networks, $G_P$ and $F_Q$, with two adversarial loss functions may not preserve the original medical image, and some reconstructed images might still contain noise. In addition to the two adversarial loss functions, the cycle-consistency loss function is required to assure that the original image and denoised image (which is generated from the generator network) have different representations, by retaining the same content and making generator networks more stable during training process. This can restrict the generator network from creating random images and ensures that the image-to-image translation is reversible. Thus, the cycle-consistency loss function $Los{s}_C(F_Q,G_P)$ is computed as

$$Los{s}_C(F_Q,G_P)={\mathbb{E}}_p\left[\parallel F_Q\left(G_P\left(p\right)\right){-p\parallel }_{L1}\right]+{\mathbb{E}}_q\left[\parallel G_P{(F_Q\left(q\right)-q\parallel }_{L1}\right]$$

(2)

where $\parallel .\parallel$ represents L1-norm. In order to ensure generated images are similar to the real images, we consider L1 loss for image similarity, which is computed as follows:

$$Los{s}_{L1}={\mathbb{E}}_p\left[\parallel G_P{\left(p\right)-p\parallel }_{L1}\right]+{\mathbb{E}}_q\left[\parallel F_Q{\left(q\right)-q\parallel }_{L1}\right]$$

(3)

Thus, the loss function of the adaptive cycle-GAN contains three components, namely, the Wasserstein adversarial loss function with two generator networks with gradient penalty, cycle-consistency loss, and L1 loss (for image similarity), and it is computed as follows:

$$Loss={\alpha }_{w}Los{s}_{WGAN}(Di{s}_P,Di{s}_Q)+{\alpha }_c^{t}Los{s}_C(F_Q,G_P)+{\alpha }_{l}Los{s}_{L1}$$

(4)

where ${\alpha }_w,{\alpha }_c^t$, and ${\alpha }_l$ are adjusted during training. The terms ${\alpha }_w$ and ${\alpha }_c^t$ represent the adaptive weighting strategies of adversarial loss and cycle-consistency loss, and they are computed as ${\alpha }_w^t={\alpha }_c^t={\alpha }_0{e}^{-\lambda t}$, where ${\alpha }_0$, $\lambda$, and t represent the initial weight, decay speed, and current epoch, respectively. Furthermore, the term ${\alpha }_l$ represents the adaptive weighting strategy of the L1 loss, and it is computed as ${\alpha }_l={\alpha }_0+{\displaystyle \frac{1}{1+\mathbb{E}\left[Los{s}_{WGAN}\right]}}$. Moreover, low adversarial loss leads to a higher L1 weight. However, the adaptive mechanism in the cycle-GAN enhances the quality of image translation by dynamically adjusting with the consideration of the adaptive cycle consistency loss parameter and dynamically tuning adversarial loss with the gradient penalty weight. In Algorithm 1, the adaptive mechanism is implemented based on adaptive cycle consistency loss and adaptive Wasserstein loss with gradient penalty. Thus, adaptive weighting helps to improve the quality of reconstructed images and the stability of the training process through the gradient penalty, which requires the Lipschitz constraint. Furthermore, during training, two generators and discriminators are updated alternately in order to maintain a balance between the two generators and discriminators. Thus, alternate training promotes the generator networks to improve in generating realistic BUS images, and discriminators can effectively distinguish real images from generated BUS images. In addition, the cycle-GAN contains the cycle-consistency loss, preserving structural integrity and improving stability during training.

Algorithm 1 Training the cycle-GAN algorithm

1: Input:            Input images, N number of iterations       Output:            Trained two Generator networks $G_p$ and $F_Q$, Discriminator networks $Di{s}_P$ and $Di{s}_Q$   2: for i = 1 to N do   3:        for each $j^{th}$ step do   4:              Generate minibatch of m samples to construct $G_P$   5:              Generate minibatch of m samples to construct $F_Q$   6:              Update Discriminator networks $Di{s}_P$ and $Di{s}_Q$   7:              Compute Wasserstein adversarial loss using (1)   8:              Compute cycle-consistency loss function using (2)   9:              Compute identity loss using (3) 10:              Update loss function of cycle-GAN using (4) 11:        end for 12: end for 13: return$CG\leftarrow$G, $G^{\prime }$, $Di{s}_P$ and $Di{s}_Q$

3.2. Ensemble-Based Multi-Stage Deep Transfer Learning Model

In deep neural networks (DNNs), increasing the number of layers can enhance the performance, but this requires more computational resources and processing time. Thus, the integration of DNNs with the transfer learning approach is an effective strategy to reduce the training costs. Through transfer learning, the learning efficacy of the new model increases by utilizing the model parameters of the previously trained model, rather than starting from scratch. In recent years, massive BUS image acquisition with data labeling is a difficult task. However, transfer learning can yield better performance, without the necessity of acquiring massive BUS images. For a sequential-based deep transfer learning model, deep convolutional layers are added, followed by a dropout with batch normalization appended to the flattened layer, for constructing an individual transfer learning model. For the efficient multi-class classification of the breast tumor identification problem, three neurons are used in the output layer of the model with softmax activation. The loss function for each individual transfer learning model is depicted in Equation (5), including the learning rate (as 0.0001) and Adam optimizer.

$$Loss=-\sum _{l=0}^M{y}_{l}log\left({\widehat{y}}_l\right)$$

(5)

where $y_l$ and ${\widehat{y}}_l$ represent the actual and predicted model output at the $l^{th}$ value. Based on the performance metrics of each model, we consider the three best transfer learning models, namely, InceptionV3, Densenet 121, and XceptionLike, with the same loss function and optimizer. InceptionV3 can achieve better performance with minimal available computational resources. Moreover, inceptionV3 incorporates auxiliary classifiers in order to stabilize the training process that leads a to fast convergence rate. Densenet 121 includes skip connections that can be utilized as a feature extractor for BUS images, and it can handle the vanishing gradient problem. Xception utilizes a depthwise separable convolution mechanism that enables the model to extract complex features from BUS images.

For better performance, we propose an ensemble of the three best sequential deep transfer learning models, namely, InceptionV3, Densenet 121, and XceptionLike. As depicted in Algorithm 2, the proposed cycle GAN with an ensemble based multi-stage deep transfer learning (EMDTL) model aggregates the predicted value from InceptionV3, Densenet 121, and XceptionLike. The transfer learning models are pre-trained on BUS images and may not provide precise results on the new data. Moreover, the proposed EMDTL model uses grouped regularization and is adaptable to the variational changes in the BUS imaging data that can effectively address the overfitting problem. However, ensemble models provide precise results rather than individual transfer learning models. Furthermore, the proposed EMDTL model can improve performance by providing adaptability and interpretability. This is because the proposed EMDTL model incorporates three different pre-trained models which are fine-tuned by different CNN architectures. Moreover, the proposed EMDTL model can handle noisy data with the incorporation of image augmentation using the cycle-GAN.

The proposed EMDTL model considers three combinations, namely, average, stacked, and hybrid approaches, in order to define the output class. For the average approach, the values of the softmax activation layer are averaged and the maximum value represents the output class label. For the stacked approach, the output of softmax layers are represented as a stacked vector, which is fed to Softmax layer in order to determine the output class label. For the hybrid approach, the last dense layer and softmax layer are summed to determine the output class label. These approaches help in optimizing the performance of the final class label. Furthermore, we consider Local Interpretable Model-Agnostic Explanations (LIME) to verify the validity of the proposed proposed EMDTL model. Algorithm 3 presents a cycle-GAN with LIME-assisted ensemble-based multi-stage deep transfer learning model (C-LEMDTL) which explores the interpretability of the transfer learning models for breast tumor identification from ultrasound images. In LIME, for each input image, multiple image perturbations are generated and trained using a surrogate model. Thus, LIME focuses on providing local approximations.

Algorithm 2 Ensemble-based multi-stage deep transfer learning model (EMDTL) algorithm

Input:            Trained augmented dataset $A_d$ using cycle-GAN, Learning Algorithms $\{L_1,L_2,\dots ,L_n\}$, $\theta$ threshold parameter       Output:   1: $i\leftarrow 1$, halt ← false   2: while (halt ≠ true) do   3:       if (i == 1) then   4:             $C_i$ = Train_Classifier($L_i,A_d$)   5:             Compute loss of $C_i$ using (5)   6:       else   7:             Learn($L_i,A_d$)   8:             $C_{i+1}\leftarrow$Transfer($C_i)\cup$ Train_Classifier($L_i,A_d$)   9:             Compute loss of $C_i$ using (5) 10:             if $|C_{i+1}|/|C_i|\ge \theta$ then 11:                   halt = true 12:                   Break 13:             end if 14:       end if 15:       i = i + 1 16: end while 17: $\overline{C}\leftarrow \mathrm{ensemble}{\left(C_i\right)}_{i\in [1,2,\dots ,n]}$

Algorithm 3 Cycle-GAN with LIME-assisted ensemble-based multi-stage deep transfer learning model (C-LEMDTL) algorithm

Input:          Trained augmented dataset $A_d$ using cycle-GAN, Learning Algorithms $\{L_1,L_2,\dots ,L_n\}$, $\theta$ threshold parameter     Output: 1: $Z\leftarrow \left\{\right\}$ 2: $A_d\leftarrow$ Train_Cycle-GAN(D) 3: for each instance $z\in A_d$ do 4:       $z^{\prime }\leftarrow$vicinity(z) 5:       f(z) ← Train_EMDTL(z) 6:       Compute the explainability of input image z $\nu \left(z\right)$ using (6) 7:       $Z\leftarrow Z\cup \nu \left(z\right)$ 8: end for

3.3. LIME for an Improved Interpretability from BUS Images

Local Interpretable Model-Agnostic Explanations (LIME) rely on over-segmenting the input image into superpixels (i.e., a set of adjacent pixels with similar features.) based on the Slic [29], Felzenszwalb [30], and Quick-Shift [31] segmentation algorithms. In this work, choosing an optimal quantity of superpixels relies on the required level of segmentation. Let Z represent the sample in the BUS image domain and binary randomized vector Z^′ = {1, 0, …, 1} be a modified perturbed sample from an explainable domain. While translating from an explainable domain Z^′ to the BUS image domain Z, pixels contained inside the superpixels are marked as ‘0’, and the other superpixels remain constant. The perturb BUS images are fed to a machine learning model F, and the predicted output $f\left(z\right)$ is recorded. Through repeating the above-mentioned process, a new BUS dataset is constructed with the set of perturbed BUS images Z and predicted output labels. Thus, the explainable model is trained on perturbed BUS image samples to train the surrogate model, and the weights assure that the surrogate model is more accurate. Moreover, higher weights are assigned to perturbed BUS images, which are similar to the input BUS image samples. The goal of LIME is to identify the surrogate model g that tries the interpretable machine learning model f, highlighting regions of the input BUS image. Mathematically, this is formulated as

$$\nu \left(z\right)=argmi{n}_{g\in \upsilon }\mathbb{L}\left(f\left(z\right),g\left(z^{\prime }\right)\right)+\alpha \psi \left(g\right)$$

(6)

where $\nu \left(z\right)$ represents the explainability of input image z, and $\upsilon$ represents a set of explainable models. Let $\alpha$ be a regularization parameter that achieves an optimal balance between interpretability and precision. The term $\psi$ represents the penalty of the surrogate model that restricts the utilization of highly complex models in order to retain interpretability. The loss function $\mathbb{L}$ uses the distance metric, which is based on the proximity of $z^{\prime }$ to z and is computed as

$$\mathbb{L}\left(f\left(z\right),g\left(z^{\prime }\right)\right)=\sum _{z,z^{\prime }}{\left(f\left(z\right)-g\left(z^{\prime }\right)\right)}^2\eta \left(z\right)$$

(7)

where $\eta \left(z\right)$ assigns higher weights to the samples that exhibit reduced perturbation in contrast to input image sample z, and whose value lie between 0 and 1. LIME uses the Gaussian kernel function to assess the perturbed image $z^{\prime }$, which relies on the close proximity to the input image sample z.

$$\eta \left(z\right)=exp\left(-{\displaystyle \frac{\Vert z-z^{\prime }{\Vert }^2)}{{\sigma }^2}}\right)$$

(8)

where $\sigma$ adjusts the dimensions of the exponential function that controls the amount of weights assigned to samples based on their proximity to z.

The main motivation behind the consideration of the explainability is to help physicians’ clinical decision making by incorporating LIME with the classification accuracy of the proposed C-LEMDTL model. Moreover, the transfer learning model mainly focuses on identifying important features of the input and not on the irrelevant background similarities. The proposed C-LEMDTL model helps in identifying a class that is vulnerable to background noise. Thus, the proposed C-LEMDTL model improves efficacy against such vulnerabilities.

4. Experimental Results

The breast ultrasound image (BUSI) dataset [2] contains images which are captured from 600 patients with 500 × 500 image dimensions. The BUSI image dataset is categorized into three classes with 487 benign, 210 malignant, and 133 normal images. During experimentation, 80% and 20% images are used to train and test the proposed C-LEMDTL model, respectively. During experimentation, K-fold cross-validation is performed to evaluate the generalization ability of the proposed C-LEMDTL model. We have conducted experimentation using a GPU-based server (which is manufactured by Tyrone Systems) with NVIDIA A100 and 80GB PCIe. Furthermore, CUDA 11.8 with cuDNN 11.8 was utilized with a batch size 16.

In this section, the performance of the proposed C-LEMDTL model and the existing methods are evaluated using a variety of metrics, such as classification reports, accuracy plots, and explainable AI techniques. For each model, the stability of the accuracy curve indicates how effectively the model performs as a classifier, and it depicts the maximum accuracy of the model. Thus, a relatively smooth curve improves classification accuracy. For the classification reports, each model efficacy is evaluated using three metrics, such as the f1-score, recall, and precision. The precision (P) represents the ratio of correctly classified outputs and the number of positively classified samples, and it is represented as $P=(T\_P/(T\_P+F\_P\left)\right)$. Moreover, the recall (R) is computed as a ratio of the number of predicted outcomes and all the evaluations of the actual class outcomes, and it is represented as $R=(T\_P+(T\_P+F\_N\left)\right)$, where $T\_P$ represents true positive, $F\_P$ represents false positive, and $F\_N$ represents false negative. Furthermore, the f1-score (F1) is computed by calculating an average of the precision and recall, and it is represented as $F1={\displaystyle \frac{2\times P\times R}{P+R}}$.

Table 1. Comparison of various pre-trained models after pre-processing.

Model	Metrics	Benign	Malignant	Normal	Accuracy
XceptionLike	Precision	0.90	0.86	0.83	0.88
	Recall	0.89	0.774	1.0	0.88
	f1-score	0.90	0.8	0.91	0.88
Inceptionv3	Precision	0.86	0.73	0.89	0.83
	Recall	0.89	0.71	0.850	0.84
	f1-score	0.88	0.72	0.87	0.83
Densenet121	Precision	0.88	0.86	0.89	0.88
	Recall	0.92	0.80	0.85	0.88
	f1-score	0.90	0.83	0.87	0.88
Resnet50	Precision	0.84	0.81	0.85	0.84
	Recall	0.89	0.70	0.85	0.84
	f1-score	0.87	0.76	0.85	0.84
MaxViT	Precision	0.47	0.68	0.54	0.38
	Recall	0.85	0.75	0.54	0.53
	f1-score	0.60	0.71	0.54	0.45

shows a comparison of the various pre-trained models for breast tumor detection. It can be observed that the XceptionLike, Inceptionv3, and DenseNet121 models have achieved the highest accuracy when compared to Resnet50 and MaxViT. To provide efficient treatment to patients, it is essential to determine the accuracy and reliability of each model using the medical image dataset. However, some models may identify and classify BUS images that are not within region of interest (RoI). As shown in Figure 4, we adopt a mask regional CNN model [32] to localize and identify the suspicious lesion in bounding boxes. In addition, Figure 4 depicts that two sample benign, malignant and normal images and masked images (named as Image 1 and Image 2) which play a major role in improving for tumor identification. In BUSI dataset, + symbol represents a reference point for lesion which helps in identifying bounding box around tumor. Later, each bounding box is fed as input to the proposed C-LEMDTL model to determine the malignancy probability value. Furthermore, there is a necessity that the proposed model must identify breast cancer regions using explainable AI techniques for the visualization of the proposed model predictions. In this work, the LIME model is utilized for the visualization of breast cancer regions.

Figure 5 shows the sample experimental results of the proposed C-LEMDTL model.

Table 2. Comparison of various models before pre-processing.

Model	Metrics	Benign	Malignant	Normal	Macro Average
XceptionLike	Precision	0.806	0.762	0.794	0.779
	Recall	0.792	0.706	0.89	0.788
Inceptionv3	Precision	0.782	0.757	0.7333	0.765
	Recall	0.771	0.602	0.750	0.759
Densenet121	Precision	0.765	0.733	0.794	0.731
	Recall	0.792	0.609	0.750	0.717
Resnet50	Precision	0.740	0.714	0.75	0.735
	Recall	0.792	0.609	0.75	0.717
MaxViT	Precision	0.472	0.681	0.542	0.384
	Recall	0.75	0.65	0.542	0.533

shows the comparison of the individual transfer learning models without pre-processing. It is observed that XceptionLike has achieved better precision, whereas Inception-V3 is the second-best classifier in terms of precision. On the other hand, MaxViT has demonstrated the worst classifier performance in terms of precision, with a value of 47.2%. Figure 6 depicts the LIME-based interpretability of the proposed model. Moreover, yellow lines highlight outline tumor boundaries that play a major role in classifying an ultrasound image as benign, malignant, or normal.

Figure 7 shows the comparison of confusion matrices for individual transfer learning models such as InceptionV3, DenseNet121, XceptionLike and the proposed C-LEMDTL model. The proposed C-LEMDTL model achieves an accuracy of 93%. The proposed C-LEMDTL model achieves better performance, while identifying more malignant tumors than the other individual transfer learning models. The proposed model reduces complexity with the consideration of the cycleGAN, which incorporates features from individual transfer learning models. The proposed C-LEMDTL model is validated using 5-fold cross validation and achieved an average of 93% accuracy on the BUSI dataset.

Table 3. Comparison of the proposed C-LEMDTL with other existing algorithms.

Method	Accuracy	Sensitivity	Precision
DeepCAD [33]	91.45	92	-
CNN [34]	65	-	-
CNNI-BCC [35]	91	88.89	-
Proposed C-LEMDTL	93	92	93

shows the comparison of the proposed C-LEMDTL model with other existing algorithms. It can be observed that the proposed C-LEMDTL model achieves the highest average accuracy. The average execution time for the proposed C-LEMDTL model with a batch size of 16 and 100 epochs is approximately 1.5–2 h. Moreover, the average execution time without a transfer learning model, with batch size of 16, and 100 epochs is approximately 3.5–4 h. This is due to the fact that transfer learning utilizes the pre-trained models, which enhances learning speed, improves stability, and reduces mode collapse.

5. Conclusions

Breast cancer medical imaging helps in early diagnosing breast cancer patients. Transfer learning with explainable artificial intelligence plays a vital role in the analysis and interpretability of breast cancer imaging. However, most of the existing studies rely on image segmentation techniques and suffer from overfitting. In this work, to avoid imbalanced data, a cycle-GAN is adopted for image augmentation, which synthesizes medical denoised images. In addition, to accurately identify augmented breast tumor images, we consider an ensemble model of pre-trained transfer learning models—Inception-v3, DenseNet121, and XceptionLike. Furthermore, LIME is adopted to provide interpretable explanations on the predictions of the proposed multi-stage transfer learning model. The experimental results have illustrated that the proposed C-LEMDTL model has achieved a 93% of accuracy and outperforms the other existing models.