Figure 1 indicates that the change in Gaussian Noise values affects to classification quality of the Deep Convolutional Neural Network, as healthy thermal images were classified as diseased. Gaussian Noise was added at the following values 0.05 and 0.02 in all thermal images of the database and the new database was saved. In addition, Deep Convolutional Neural Network Inception MV4 was set to a learning rate of 0.0001, SGDM optimization method and training 10 epochs see Figure 2. As for the second stage, Deep Convolutional Neural Network was trained using the same settings, but without adding Gaussian Noise. The third stage was the training of Deep Convolutional Neural Network on the same settings, but by adding actual thermal images with added Gaussian Noise.
The following graph shows the amount of change in Noise compared to the accuracy of classification in each of DMR IR and DMR IR databases with Noise and with Gaussian Noise. Where the results indicate that the database without Noise has best solution for classification. As for database with Noise in any of the thermal images, filters are added during image processing and before entering Deep Convolutional Neural Network. Figure 3 is clear from the comparison between thermal image 1, thermal image 2, and thermal image 3 constitutes the difference between the addition of Gaussian Noise and the actual thermal image. Moreover, it is clear from Figure 4 gradient in addition to Gaussian Noise and the extent of its impact on the accuracy of classification. Table 2 shows that the analysis reveals a mean noise value, a standard deviation of noise, and a noise value in dB from Figure 4. These parameters impact on the classification accuracy in thermal images. Fourth stage: thermal image processing filters built before being inserted into Deep Convolutional Neural Network. Detection accuracy was calculated and found to highlight the effect of filters on the classification using Deep Convolutional Neural Network. The results indicate that the accuracy of training and verification in the database containing Gaussian Noise is less than the detection accuracy of the database without Gaussian Noise. Adjusting Gaussian Noise on the mean and the variance has different values from 0.01 to 0.09. The results indicate that greater Noise has a higher false classification, as the classification of thermal images reached 90% of false classification. Furthermore, during training, Deep Convolutional Neural Network Inception MV4 due to combining of actual thermal images with Noise thermal images with mean and variance values of 0.05 and 0.02 respectively and storing new thermal images in a single database. The results indicate that change in the accuracy of false detection, which is a very small mean value of 0.01 despite the change in variance with high values as shown in Figure 5.
As for the second stage, the database of thermal images was used to which only Gaussian Noise was added without adding actual thermal images but storing modern thermal images as a new database. Therefore, the results show that the database has a slight impact on the accuracy of detection, as shown in Figure 6. Through the graph, they show a significant change in the mean of 0.01 and the variance of 0.01 compared to other values that maintained an accuracy of 100%. In the third stage, the MV4 Deep Convolutional Neural Network was verified on the main database consisting of 1800 thermal images without adding the Gaussian Noise. The results indicate a detection accuracy of 100%.
Table 3 shows comparison between three thermal image databases after training Deep Convolutional Neural Network Inception MV4 after setting learning rate to 0.0001, training 10 epochs and using SGDM optimization method. Whereas best results indicated by table indicate that database does not contain Gaussian Noise is better than database that contains Gaussian Noise. Furthermore, average accuracy for database containing Gaussian Noise with actual thermal images and database containing only Gaussian Noise and the database without Gaussian Noise were 99.964%, 98.8% and 99.974% respectively.
Table 4 shows addition of salt and pepper noise while adjusting variables from 0.1 to 0.3, with measuring accuracy several times and calculating average accuracy. The results show that deep convolutional neural network MV4 model was affected in accuracy average between 0.2 and 0.3 from 100% to 51.58%, without adding a filter in the pre-processing. Table 5 shows Variance of multiplicative noise, with variables adjusted from 0.2 to 0.8. With increase in noise, the classification accuracy was affected only in values 0.7, 0.8, which were 89% and 43%, respectively. On the other hand, when using speckle noise while adjusting variables from 0.2 to 0.8, the results show that classification accuracy decreased in value 0.8, however; it maintained its performance in remaining values as shown in Table 6. Table 7 indicates evaluation parameters such as accuracy, sensitivity, Specificity …etc.
In conclusion, it is clear from the four tables that salt noise is the most influential on the classification accuracy of deep convolutional neural network MV4 without using filters during pre-processing. As for the use of filters in the pre-processing, it is clear that classification accuracy reached 97.999%. Table 8 focuses on five different values of Poisson noise setting (0.2 to 0.8) and examines the performance at various SNR (13.98 dB to 1.94 dB). The objective is to analyse how changes in SNR and affect the accuracy of the thermal image’s classification.
Table 5 presents the average detection accuracy for various values in the Variance of multiplicative noise. In addition, it displays accuracies corresponding to different settings ranging from 0.2 to 0.8. On the other hand, in setting 0.2, the accuracy for all five measurements is a perfect 100%. Also, as variance increases to 0.3 and 0.4, the accuracy remains consistently high at 100%. However, when the variance reaches 0.5 and 0.6, we observe a slight decrease in accuracy, dropping to 99.9958% and 99.9026% respectively. Subsequently, in settings 0.7 and 0.8, accuracy notably decreases further to 89.6558% and 43.86% respectively. Finally, results indicate a trend where higher values of variance in multiplicative noise negatively impact the detection accuracy, with a significant drop observed beyond a variance of 0.6.
In table 6 showing average detection accuracy of different values in speckle noise, various settings were tested with noise levels ranging from 0.2 to 0.8. The result indicates that at lower noise levels such as 0.2 and 0.3, the accuracy of detection was consistently at 100%. Moreover, as noise level increased, there were slight decreases in accuracy with most significant drop occurring at a noise level of 0.8. An accuracy 1 maintained a high accuracy rate across all noise levels, with a slight decrease to 95.4% at a noise level of 0.7. Accuracy 2 also showed high accuracy rates but experienced a significant drop to 44.5% at a noise level of 0.8. Accuracy 3 and Accuracy 4 had fluctuations in accuracy, with Accuracy 4 showing a noticeable decrease to 69% at a noise level of 0.6. Overall, the average accuracy across all settings decreased as the noise level increased, with the highest average accuracy of 100% observed at lower noise levels of 0.2 and 0.3. Also, average accuracy dropped to 62.114% at a noise level of 0.8, indicating a significant impact of higher noise levels on detection accuracy. Finaly, these results suggest that while the detection accuracy remains high at lower noise levels, there is a notable decline in performance as the noise level increases, highlighting the importance of noise reduction techniques in improving detection accuracy in speckle noise environments.
The experiment was conducted by measuring the accuracy of the deep learning inception mv4 at different values of Poisson noise. The accuracy measurements were recorded for each SNR level, and the average accuracy was calculated across all measurements. The settings were denoted by SRN (Signal to Noise Ratio). The obtained data is presented in Table 6. From Table 6, at accuracy 1, it can be observed that at Poisson noise of 0.2, the accuracy is recorded as 100%, indicating perfect performance. However, as the Poisson noise increases, the accuracy decreases. At Poisson noise of 0.3 in Figure 7, the accuracy drops to 99.9%, suggesting a slight degradation in performance. As the Poisson noise increases further, the accuracy gradually improves, reaching 100% at Poisson noise of 0.6. Accuracy 2, Accuracy 3, Accuracy 4, Accuracy 5: These accuracy measurements follow a similar pattern to Accuracy 1. At lower Poisson noise values (0.2 to 0.4), the accuracy is considerably lower, indicating reduced performance. However, as the Poisson noise increases, the accuracy improves and eventually reaches 100% at a Poisson noise of 0.6. Moreover, the average accuracy across all measurements shows a consistent pattern. It is high (99.9% or 100%) at Poisson noise of 0.4 and above, indicating excellent overall performance. Finally, when the Poisson noise numbers used in the experiment are even decimal numbers, the corresponding results consistently exhibit a percentage of 100%. In contrast, when the noise numbers are odd decimals, there is a slight decrease in the observed percentage, specifically by 0.01%. Table 9 showcases the confusion matrices for three different deep learning models: Inception v3, Inception v4, and Inception mv4, each evaluated for their performance in breast cancer detection. Also, these matrices encapsulate the fundamental metrics crucial for assessing the models' classification accuracy.