The segmentation and classification models are tested on several datasets. For segmentation, BUSIS[24] and DDSM [25] datasets are utilized. For classification, Histopathology [26] and BreakHis [27, 28] datasets are utilized. MATLAB-2023B, Google Colab, and Jupyter Notebook with Windows OS and RTX 3070 graphic card do the studies. This research tests the suggested technique with two experiments. The recommended techniques are assessed using sensitivity, accuracy specificity, and F1 score. We abbreviate these as Sn, Acc, SP, and f1 respectively. The dataset is described in Table 4 These BC detection databases are public.
Table 4
Dataset detail used in research.
Ref # | Years | Datasets | Classes | No of Images |
[24] | 2021 | BUSIS | 2 | 1,578 |
[25] | 2021 | DDSM | 2 | 13,140 |
[26] | 2018 | Histopathology | 2 | 271,404 |
[27] | 2019 | BreakHis | 2 | 16,143 |
[28] | 2020 | BreakHis | 8 | 196,868 |
Table 4 displays the number of input images that are applied for training and testing. The detail of the publicly available dataset is described as follows: BreakHis dataset contains two classes of BC such as malignant and benign with 16,143 images. BreakHis dataset contains 196,868 images with eight classes including AD, FB, PT, TA, DC, LC, MC, and PC lesions. The DDSM dataset contains 2 classes such as malignant and benign with 13,140 images. The BUSIS contains 1,578 images with 2 classes. The Histopathology contains 271,404 images for 2 classes such as malignant and benign. One experiment is performed on segmentation of the breast lesions and two models are proposed for classification of the breast lesions.
4.1. Experiment#1: Segmentation of the Breast Cancer (BC)
The proposed segmentation model's performance is computed using mean and weighted IoU as M-IoU and W-IoU, global (G-Accuracy) as G-Acc and mean (M-Accuracy) as M-Acc, mean BF score as M-BF-Score from BUSIS and DDSM datasets in Tables 5–6.
Table 5
Performance of the segmentation model using the BUSIS dataset.
G-Acc | 0.991 |
M-IoU | 0.990 |
M-Acc | 0.989 |
W-IoU | 0.988 |
M-BF-Score | 0.979 |
Table 5 shows the suggested segmentation model's 0.989 and 0.991 BUSIS M- and G-Accuracy. BUSIS and DDSM breast cancer segmentation findings are shown in Figs. 6 and 7.
Using the BUSIS dataset, segmented BC is overlaid on original input pictures to highlight breast cancer locations in Fig. 6.
Table 6
Performance of the segmentation model using the DDSM dataset.
G-Acc | 0.989 |
M-IoU | 0. 987 |
M-Acc | 0. 986 |
W-IoU | 0. 989 |
M-BF-Score | 0. 987 |
Table 6 summarizes the results of the segmentation proposed earlier; the model achieved M-Acc and G-Acc of 0.986 and 0.989 on the DDSM dataset. Figure 7 presents segmented breast cancer and maps it to the source images to demonstrate diseased breast areas using the DDSM dataset.
The achieved segmentation results are compared to existing methods as given in Table 7.
Table 7
Proposed model for segmentation result compared with existing work.
Ref # | Year | Datasets | Result (ACC) |
[29] | 2020 | CBIS-DDSM | 84.40% |
[30] | 2021 | 91.99% |
[31] | 2022 | 92.86% |
[32] | 2023 | 86.71% |
[18] | 2024 | 92.00% |
Proposed Method | 98.00% |
[33] | 2021 | BUSIS | 94.12% |
[34] | 2022 | 89.73% |
[35] | 2023 | 96.52%,93.18% |
[36] | 2024 | 96.99% |
Proposed Method | 99% |
For automatic breast segmentation, use Atreus neural semantic segmentation. This DeepLabv3 + investigation discovered malignancies on CBIS-DDSM and BUSIS datasets 98% of the time. The intended VGG16 network is 0.844 CBIS-DDSM accurate[29]. The two-view classifier showed 0.9199 ± 0.0623 accuracy in 5-fold cross-validation, identifying malignant and non-cancerous breast images using one model and no additional data[30]. INbreast scored 96.34% for Connected-SegNets, CBIS-DDSM 92.86%, and private 92.25% [31]. The method achieved 86.71% CBIS-DDSM dataset accuracy[32]. Transfer learning using the VGG16 model was 92–95% accurate. VGG16 transfer learning benefits CBIS-DDSM and UPMC[18]. The public BUS dataset gave the SHA-MTL model 94.12% accuracy[33]. CNN model MIAS accuracy was 96.55%. The DDSM dataset showed 90.68%. INbreast showed 91.28% [34]. ShuffleNet-ResNet finds 99.17% abnormalities and 98.00% malignancies in mini-DDSM and 96.52% and 93.18% in BUSI datasets[35]. The triple decoder + multi-attention model was 96.99% accurate on BUSI and 97.69% on UDIAT. Jaccard index testing is 83.40% in UDIAT and 82.31% in BUSI [36].
4.2. Experiment#2: XAI Model Classification of BC
The XAI model is used for the classification of BC. The BC is classified into two classes: malignant and benign. Table 8 demonstrates the BreakHis dataset's results using our approach to the classification problem. The training and testing outcomes of the proposed method are provided in Fig. 8. Accuracy as ACC, F1 Score as F1, Recall as Re, and Precision as Pr.
The classification results on BreakHis are presented in the form of a confusion matrix in Fig. 9.
Table 8
Proposed classification result on BreakHis Dataset with magnification factors.
Magnification Factor | ACC | F1 | Re | Pr |
40X | 1.00 | 1.00 | 1.00 | 1.00 |
100X | 1.00 | 1.00 | 1.00 | 1.00 |
200X | 1.00 | 1.00 | 1.00 | 1.00 |
400X | 1.00 | 1.00 | 1.00 | 1.00 |
Table 8 shows 100% accuracy in the BreakHis dataset for classification. Table 9 compares the proposed classification model to current studies.
In Fig. 10, computed results on eight classes present the ratio of true positive/negative and false positive/negative. The training and validation accuracy with the number of epochs is presented graphically where a pink line for training and a green line for validation. The quantitative results on eight classes using the BreakHis dataset with a 40x magnification factor are given in Table 9 and the 100x magnification factor is given in Table 10.
Table 9
Classification results on BreakHis with 40x magnification factor
Classes | Precision% | Recall% | F1-score% |
AD | 1.00 | 0.91 | 0.95 |
DC | 1.00 | 1.00 | 1.00 |
FB | 0.86 | 1.00 | 1.00 |
LC | 1.00 | 1.00 | 1.00 |
MC | 0.92 | 1.00 | 0.96 |
PC | 1.00 | 0.81 | 1.00 |
PT | 1.00 | 0.81 | 0.89 |
TA | 0.94 | 1.00 | 0.97 |
Accuracy | 0.96 |
Misclassification rate | 0.03 |
Macro-F1 | 0.96 |
Weighted-F1 | 0.96 |
Table 9, provides the classification outcomes on 40x magnification factor on eight classes such as AD, FB, PT, TA, DC, LC, MC, and PC. In which 96.27% accuracy and 0 misclassification rate is achieved.
Table 10
Classification results on BreakHis with 100x magnification factor
Classes | Precision% | Recall% | F1-score% |
AD | 1.00 | 0.98 | 0.99 |
DC | 1.00 | 1.00 | 1.00 |
FB | 1.00 | 0.51 | 0.68 |
LC | 0.89 | 1.00 | 0.94 |
MC | 0.96 | 0.70 | 0.81 |
PC | 1.00 | 1.00 | 1.00 |
PT | 0.44 | 1.00 | 0.61 |
TA | 1.00 | 0.95 | 0.97 |
Accuracy | 0.85 |
Misclassification rate | 0.14 |
Macro-F1 | 0.87 |
Weighted-F1 | 0.85 |
Table 10, provides the classification outcomes on 100x magnification factor on eight classes such as AD, FB, PT, TA, DC, LC, MC, and PC. In which 85.03% accuracy and 0 misclassification rate is achieved.
In Fig. 11, computed results on eight classes present the ratio of true positive/negative and false positive/negative. The training and validation accuracy with the number of epochs is presented graphically where a pink line for training and a green line for validation. The quantitative results on eight classes using the BreakHis dataset with a 40x magnification factor are given in Table 11 and the 100x magnification factor is given in Table 12.
Table 11
Classification results on BreakHis with 200x magnification factor
Classes | Precision% | Recall% | F1-score% |
AD | 0.97 | 1.00 | 0.98 |
DC | 1.00 | 1.00 | 1.00 |
FB | 1.00 | 0.95 | 0.97 |
LC | 1.00 | 1.00 | 1.00 |
MC | 1.00 | 1.00 | 1.00 |
PC | 1.00 | 1.00 | 1.00 |
PT | 0.95 | 1.00 | 0.97 |
TA | 1.00 | 0.97 | 0.98 |
Accuracy | 0.99 |
Misclassification rate | 0.00 |
Macro-F1 | 0.99 |
Weighted-F1 | 0.99 |
Table 11, provides the classification outcomes on 100x magnification factor on eight classes such as AD, FB, PT, TA, DC, LC, MC, and PC. In which 99.10% accuracy and 0 misclassification rate is achieved.
Table 12
Classification results on BreakHis with 400x magnification factor
Classes | Precision% | Recall% | F1-score% |
AD | 0.91 | 1.00 | 0.95 |
DC | 1.00 | 1.00 | 1.00 |
FB | 1.00 | 0.98 | 0.99 |
LC | 1.00 | 0.98 | 0.99 |
MC | 1.00 | 0.92 | 0.96 |
PC | 1.00 | 1.00 | 1.00 |
PT | 1.00 | 1.00 | 0.99 |
TA | 1.00 | 1.00 | 1.00 |
Accuracy | 0.98 |
Misclassification rate | 0.01 |
Macro-F1 | 0.98 |
Weighted-F1 | 0.98 |
Table 12, provides the classification outcomes on 100x magnification factor on eight classes such as AD, FB, PT, TA, DC, LC, MC, and PC. In which 98.62% accuracy and 0 misclassification rate is achieved.
4.3. Experiment#3: Classification of the BC using the QNN Model
The QNN model is used for the classification of BC. The BC is classified into two classes: malignant and benign. Figure 10 shows the classification result on the publicly available BreakHis dataset.
Figure 12, presents the classification results on different magnification levels such as 40x, 100x, and 200x. The performance of classification is measured in terms of the confusion matrix and training and validation accuracy for the number of epochs. In the confusion matrix, 0 denotes the benign class and 1 is the malignant class. The quantitative assessment in terms of ACC, F1, Re, and Pr.
Table 13
classification result on BreakHis Dataset with magnification factors.
Magnification Factor | ACC% | F1 | Re | Pr |
40x | 1.00 | 1.00 | 1.00 | 1.00 |
100x | 0.97 | 0.97 | 1.00 | 0.95 |
200x | 0.97 | 0.97 | 1.00 | 0.94 |
In Table 13, the achieved results in terms of accuracy based on the binary classification are 100% on 40x, 97.92% on 100x, and 200x. The classification results are computed on eight classes of breast cancer and are presented in Fig. 11.
In Fig. 13, computed results on eight classes present the ratio of true positive/negative and false positive/negative. The training and validation accuracy with the number of epochs is presented graphically where a pink line for training and a green line for validation. The quantitative results on eight classes using the BreakHis dataset with a 40x magnification factor are given in Table 14.
Table 14
Classification results on BreakHis with 40x magnification factor
Classes | Precision% | Recall% | F1-score% |
AD | 1.00 | 1.00 | 1.00 |
FB | 1.00 | 1.00 | 1.00 |
PT | 1.00 | 1.00 | 1.00 |
TA | 1.00 | 1.00 | 1.00 |
DC | 1.00 | 1.00 | 1.00 |
LC | 1.00 | 1.00 | 1.00 |
MC | 1.00 | 1.00 | 1.00 |
PC | 1.00 | 1.00 | 1.00 |
Accuracy | 1.00 |
Misclassification rate | 0.00 |
Macro-F1 | 1.00 |
Weighted-F1 | 1.00 |
Table 14, provides the classification outcomes on 40x magnification factor on eight classes such as AD, FB, PT, TA, DC, LC, MC, and PC. In which 100% accuracy and 0 misclassification rate is achieved.
Table 15
Classification results on BreakHis with 100x magnification factor
Class name | Precision | Recall | F1-score |
AD | 1.00 | 1.00 | 1.00 |
FB | 1.00 | 1.00 | 1.00 |
PT | 1.00 | 1.00 | 1.00 |
TA | 1.00 | 1.00 | 1.00 |
DC | 1.00 | 1.00 | 1.00 |
LC | 1.00 | 1.00 | 1.00 |
MC | 1.00 | 1.00 | 1.00 |
PC | 1.00 | 1.00 | 1.00 |
Accuracy | 1.00 |
Misclassification rate | 0.00 |
Macro-F1 | 1.00 |
Weighted-F1 | 1.00 |
Table 15 provides the classification outcomes on the 100x magnification factor. The achieved outcomes are compared to the existing methods as given in Table 16.
Table 16
Proposed model for classification result compared with existing work
Ref # | Year | Datasets | ACC% |
[37] | 2021 | BreakHis 2 classes | 96.75 |
[38] | 2022 | 94.67 |
[39] | 2022 | 96.35 |
[40] | 2023 | 96.30 |
[41] | 2024 | 99.16 |
[19] | 2024 | 99.76 |
Proposed Method with XAI | 96.87 |
Proposed Method with QNN | 1.00 |
[42] | 2021 | BreakHis 8 classes | 95.00 |
[43] | 2023 | 89.00 |
[41] | 2024 | 98.27 |
Proposed Method with XAI | 98.62 |
Proposed Method with QNN | 1.00 |
XAI and QCNN models classify BC with 96.87% and 100% accuracy on BreakHis with 2 and 8 classes, respectively, according to Table 16. Model accuracy is 96.75%, 96.7%, 95.78%, and 93.86% for benign and malignant binary classification [37]. Classify BreakHis with pretrained models. The highest classifier is VGG16 with a precision of 92.60% and an f1- score of 85.21%. The model achieved an accuracy of 94.67% and a recall of 80.52% [38]. The proposed model has a maximum accuracy of 96.35%. The network architectures tested include VGG16, ResNet50, and a proposed model [39]. On the BreakHis dataset, pretrained models had 92%, 87%, 90%, 79%, and 92% accuracy. Data features were retrieved using ResNet 50. To 96.3%, accuracy has substantially increased [40]. AlexNet convolutional neural network diagnoses breast cancer using BreakHis dataset features. The proposed method has 99.36% AUC, 95% accuracy, as well as 97% sensitivity with 90% specificity [42]. Deep learning cancer classifiers VGG, ResNet, Xception, Inception. Top performance Xception is 0.9 F1 and 89% accurate. Inception and ResNet are 87% accurate [43]. The Swin-Transformer V2 architecture classified eight-class BC histopathological images with multiple labels with 98.27%, 97.95%, 98.97%, and 99.16% accuracy on Break-His [41]. Rapid Tri-Net with Aquila Optimization (Rapid Tri-Net) on BreakHis and BACH datasets achieved 99.79%, 99.8%, 99.73%, and 99.76% accuracy at 40x, 100x, 200x, and 400x [20].