Fig 1.
Few samples of digitized mammography images collected from the MIAS dataset.
(a-b) are benign tumors, (c-d) malignant tumors, (e-f) sample of extracted ROIs of benign tumors, (g-h) extracted ROIs of malignant tumors.
Fig 2.
Randomly selected samples of mammographic images used after preprocessing and enhancement, starting with original and augmented images (a) Original raw mammogram image with breast masses, (b) Vertical flipped, (c) Horizontally flipped, (d) Rotation +45, (e) Rotation -45, and (f) Cropping.
Table 1.
Different ques of augmenting data with invariance parameters.
Fig 3.
The layered framework of the designed CNN-based model (ConvNet) consists of 3 × 3 Con_2D convolutional layers, followed by the max-pooling layers and softmax classifier for classifying mammography abnormalities as malignant or benign.
Fig 4.
The implementation of designed ConvNet architecture containing convolution2D, ReLu, max-pooling, batch normalization and softmax layer for the classification of breast masses.
Table 2.
Hyperparameters configurations of proposed architecture.
Table 3.
Performance comparison of proposed model with pre-trained models on Private dataset.
Fig 5.
(a) Represent the training and testing accuracy of the proposed model against the training step using a Private dataset. The training accuracy scaled up on each epoch and after the 90th epoch was 0.98 which other models outperform. (b) Represent training and testing cross-entropy loss of the proposed model against the training step using Private dataset. The proposed model generates reduced false-negative results compared to existing DCNN models. The training cross-entropy loss function is declining continuously on every epoch and reduced to a minimum equal to 0.068 at the last epoch.
Fig 6.
(a) Depicts the suggested system’s accuracy and cross-entropy loss throughout the training phase. The ascending curves indicate the proposed system’s accuracy, while the descending curves indicate the proposed system’s loss value on the training and testing MIAS dataset. The training and testing curves are nearly identical, suggesting that the model was appropriately trained. (b) Training and testing precision, F-score and sensitivity of proposed model using MIAS dataset.
Table 4.
Performance comparison of proposed model with pre-trained models on MIAS dataset.
Fig 7.
(a) Represent the training and testing area under the curve of the proposed model using Private dataset and MIAS datasets. The graph shows that the proposed model has attained a maximum AUC of 0.99. The experimental findings indicate that the suggested model yields a small number of false positives and high true positives. (b) Training and testing precision, F-score and sensitivity of proposed model using Private dataset. The proposed model has performed well by attaining an excellent sensitivity of 0.99.
Fig 8.
(a) The proposed architecture improves the average accuracy rate, AUC, precision, and sensitivity on Private dataset than MIAS dataset. The proposed model generates reduced false-negative results compared to existing DCNN models. (b) depicts the training accuracy, loss, sensitivity, AUC and precision of proposed ConvNet architecture and DCNN models on MIAS dataset. The proposed model shows promising performance compared to state-of-the-art standard models.
Table 5.
Comparison of proposed approach with state-of-the-art conventional schemes in terms of accuracy rate.
The experimental findings demonstrate the effectiveness of the suggested approach, with an overall training 0.98 accuracy.