An Accurate Diagnosis and Classification of Breast Mammogram Using Transfer Learning in Deep Convolutional Neural Network

Utilizing computer-aided detection (CAD) technologies for early mammogram detection may enhance treatment results and increase survival rates for breast cancer patients. Conventional CAD software depends on manually identified features, which have limitations. Hand-crafted features are sometimes peculiar to a particular field and the process of designing them may be laborious, challenging, and not easily transferable.

CAD methods enhance early breast cancer identification and treatment results; nevertheless, conventional CAD depends on human-designed features that are frequently arduous, domain-specific, and challenging to transfer between datasets. Progress in deep learning, specifically the Convolutional Neural Networks (CNNs), provides a more effective method for the automatic extraction of information from medical images.

The development of a cancer tumor is attributed to the unregulated growth and division of cells, which subsequently invade and spread into the surrounding tissues within the human body [1]. The Convolutional Neural Network (CNN) is applied as another approach for extraction of features from the image directly [2]. CNN has excelled in several picture classification challenges. CNN has been used in medical picture categorization using the three primary methods. 1) Training a CNN from the beginning 2) Use of feature extraction 3) Adjusting a pre-trained model on images [3].

One effective transfer learning approach involves pre-training network parameters on source data, applying these parameters to the target domain, and then fine-tuning the network for improved performance. A methodology for locating and organizing multi-class breast cancer based on transfer learning has been suggested and put into practice in this scenario. The suggested model has two primary components [4]. The first component has six primary phases intended to improve the level of quality of the breast images. And second component is a hybrid deep learning algorithm that includes pre-trained CNN like VGG16 is used to transmit its acquired features to the mammogram classification [5].

Nonetheless, deficiencies persist in the diagnosis of multi-class breast cancer, especially with transfer learning. This research introduces an innovative methodology that improves image quality and utilizes a pre-trained CNN Modified Inception v3 for enhanced multi-class breast cancer classification.

The approach for detecting and classifying breast cancer has two primary components. The first component is for preprocessing the picture, while the subsequent one is employed for transferring the CNN parameters [14]. Figure 1 displays the intricate structure of the proposed work.

1.png

Figure 1. Proposed workflow

3.1 Image preprocessing

Image preprocessing is crucial in mammogram detection to improve image quality and ready it for precise analysis by CAD systems. Here are some essential stages.

2.png

Figure 2. Levels of data preprocessing

Before the learning process is initiated, segmentation methods can be used to clearly auto detect the tumor areas. To cut down computing time, this is done. The enhancement of image visibility and the improvement of segmentation accuracy can be achieved by reducing background noise, equalization of histograms, and morphological analysis techniques prior to segmentation. Figure 2 illustrates that data preparation involves seven steps.

3.1.1 Noise removal

Mammograms can be affected by variety of noises, such as gaussian, speckle, salt-and-pepper noise. This noise may diminish the efficacy of classification methods. Using a 2D median filter may efficiently eliminate noise while retaining crucial information for categorization. Figure 3 displays the picture after the implementation of a 2D median filter.

3.png

Figure 3. Normal vs noise removed image

Let I_med(x,y) represent the intensity of the same pixel after applying a 2D median filter. The 2D median filter analyzes a square or rectangular neighborhood positioned in the middle of each pixel. Define N_x,yas the nXn neighborhood centered at pixel (x,y). The median value, denoted as I_med(x,y), is calculated as the median of pixel intensities in the neighborhood N_xy. Mathematically, the computation of I_med(x,y) can be represented in Eq. (1) as follows:

$I_{\text {med }(x, y)}=\operatorname{median}\left(N_{x y}\right)$  (1)

Algorithm 1: 2D median filter for noise removal

1) Input: load the mammogram image as a grayscale image I_(x,y)

2) Define Kernel Size: Determine the size of the square or rectangular neighborhood (kernel) around each pixel. Common choices include 3×3, 5×5, or 7×7 kernels, but the size can vary depending on the specific application and desired filtering effect.

3) Initialize Output Image: Generate a blank output image matrix with dimensions identical to the input picture to store the filtered image.

4) Iteration: Perform an iteration over every pixel’s intensity (x, y) in the input image.

5) Apply Median Filter:

i. For each pixel location (x, y), define a neighborhood N_xy centered at that pixel, with a size determined by the kernel size.

ii. Collect the intensity of pixels in the surrounding area N_xy.

iii. Sort the intensity of pixels in ascending order.

iv. Determine the sorted pixel's median value. If the neighborhood's pixel count is even, take the average of the two intermediate values.

v. The computed median value is assigned to the equivalent pixel location in the output image.

6) Output: The resulting output image is the filtered image obtained after applying the 2D median filter to the input mammogram image.

3.1.2 Contrast enhancement

Histogram equalization (HE) is a prevalent method for improving picture contrast, often used in image processing, including mammograms. It adjusts the intensity levels of the picture to cover the whole dynamic range, improving contrast and making the image more visually attractive and simpler to analyze [15]. Figure 4 displays the original picture beside the image with contrast enhancement achieved using histogram equalization.

4.png

Figure 4. Contrast enhancement using HE

Algorithm 2: Histogram equalization for Contrast Enhancement

Input: The variable I(x, y) represents a grayscale mammogram image. The pixel values in this image range from 0 to L-1. The variable L represents the number of intensity levels, which is typically set to 256 bits.

Output: g(x, y)-Contrast-enhanced mammogram image

1) Calculate Histogram H(k): Histogram of image I, where, H(k) represents the number of pixels with intensity k (0≤k≤L-1)

2) Normalize Histogram n(k): The normalized histogram represented in Eq. (2), obtained by dividing H(k) by the overall pixel (M×N): Where M and N are the image height and width.

$n(k)=\frac{H(k)}{M \times N}$   （2）

3) Calculate Cumulative Distribution Function (CDF):

p(k): CDF of the normalized histogram, representing the probability of finding a pixel with intensity less than or equal to k, shown in Eq. (3):

$p(k)=\sum_i^k n(i)$    (3)

4) Map Intensities To New Values: g(x, y) equalized image with enhanced contrast, represented in Eq. (4):

$g(x, y)=L-1 * p(I(x, y))$   （4）

3.1.3 Morphological analysis

Morphological analysis is a crucial step in eliminating non-breast areas prior to segmentation to ensure the accuracy of the findings. By applying the structural element (SE) to the input image, the necessary structures may be extracted during morphological operations. The resulting image obtained from this process retains the exact identical dimensions of the original input image displayed in Figure 5. The value of each pixel is based on the value of the pixel next to it and the value of the same pixel in the input picture.

5.png

Figure 5. Original vs morphological mammogram image

Algorithm 3: Morphology Analysis

1) Set Representation: Every pixel in the mammogram is represented as a binary value (0 or 1), with 0 for black and 1 for white. The entire image is considered as a binary set X, where each element corresponds to a pixel.

2) Structuring Element: This is a small binary image defining the shape used for morphological operations. For instance, a 3×3 square with all ones represents a square SE. The size and shape of the SE significantly impact the outcome of operations.

3) Erosion: Erosion shrinks objects in the image by removing pixels that don't touch the SE when placed at every pixel location in X. Mathematically, erosion of set X by SE B (denoted as X⦵B) is defined in Eq. (5), where $B_X$ is the shifted version of B placed at pixel x.

$X \ominus B=\left\{X \in X \mid B_X \subseteq X\right\}$  （5）

4) Dilation: Dilation expands objects by adding pixels that touch the SE when placed at every pixel location in X. Mathematically, dilation of set X by SE B (denoted as $\mathrm{X} \oplus \mathrm{B}$ ) is defined in Eq. (6):

$X \oplus B=\left\{X \mid B_X \cap X \neq \emptyset\right\}$   (6)

5) Opening and Closing: Opening combines erosion and dilation to remove small objects and smooth edges, represented in Eq. (7):

$I O=\operatorname{In} p \ominus S E \oplus S E$  (7)

Closing combines dilation and erosion to fill small holes and smooth edges, represented in Eq. (8):

$I C=I n p \oplus S E \ominus S E$ (8)

Opening and closing are defined using erosion and dilation iteratively.

3.2 Segmentation

By applying a threshold-based segmentation approach for automated region extraction, the calculation time may be minimized and the evaluation can be concentrated on the area most impacted by cancer. This work utilizes a global threshold for segmenting mammogram images.

6.png

Figure 6. Original vs segmentation

Global thresholding is a common and straightforward technique used to segment images, such as mammograms. This approach involves applying a single threshold value to the whole picture to distinguish between foreground (breast tissue) and background. Figure 6 displays the segmented picture derived from the original.

Algorithm 4: Global Threshold based Segmentation

1) Select Threshold Value T_h: Which separates foreground and background in the grayscale image. This can be done manually or automatically using techniques like Otsu's method.

2) Apply the threshold value T_h: To the grayscale image. Here M(x,y) shown in Eq. (9) is the binary mask indicating segmented regions.

$M(x, y)=f(x)=\left\{\begin{array}{l}1 \text { if } I(x, y)>T_h \\0 \text { Otherwise }\end{array}\right.$    (9)

I(x,y) is the intensity of the grayscale image at pixel (x,y). T_h is the threshold value. The resulting binary mask M(x,y) contains foreground regions (breast tissue) represented by pixel value 1, and background regions represented by pixel value.

3.3 Data augmentation

Deep learning techniques perform more well with extensive datasets. Data augmentation is an extensively used method to expand the dataset, which aids in mitigating overfitting when working with a tiny quantity of data during training. In this work, both training and testing data may be expanded by a series of changes.

The Data augmentation algorithm (DAA) is used to enhance the amount of input data. Figure 7 the segmented pictures undergo clockwise rotation by 90°, 180°, 270°, and 360°. Subsequently, each rotating picture is vertically mirrored. An input picture will generate eight images using this method.

7.png

Figure 7. Image augmentation using DAA

Algorithm 5: Data Augmentation

1) Random Rotation: Rotate the image and corresponding label by a random angle within a specified range represented in Eqs. (10) and (11):

Image rotation

Rotated Image $=$ rotate(I, angle) （10）

Label Rotation

Rotated Label $=$ rotate $(\mathrm{L}$, angle $)$ (11)

2) Random Shear: Apply a random shear transformation to the image and label. Calculate the shear Transformation Matrix shown in Eq. (12):

Shear Matrix $=\left[\begin{array}{ll}1 & \text { shear factor } \\ 0 & 1\end{array}\right]$     (12)

Image Shear using Eq. (13):

$\text { Sheared Image }  =\text { affine transform(I, shear matrix) }$  （13）

Label Shear using Eq. (14):

Sheared label = affine transform(L, shear matrix)  （14）

3) Random Zoom: Randomly zoom in or out of the image and label. Calculate zoomed shape using Eq. (15):

$\text { Zoomed Shape }=\text { (int (h X zoom - factor), } \text { int(w X Zoom - factor) })$   (15)

Image_Zoom using Eq. (16):

Zoomed Image $=$ zoom $(\mathrm{I}$, zoom - factor $)$   (16)

Label_Zoom using Eq. (17):

$\text{Zoomed }\!\!~\!\!\text{ Label}=\text{zoom}\left( \text{L},\text{ }\!\!~\!\!\text{ zoom}-\text{factor} \right)$   (17)

4) Random Horizontal flip: Represented in Eqs. (18) and (19). Optionally perform a horizontal flip with a probability of 0.5.

Flipped Image $=$ flip - left $-\operatorname{right}(\mathrm{I})$  (18)

Flipped Label $=$ flip - left $-\operatorname{right}(\mathrm{L})$ (19)

5.1 Dataset preparation

The proposed approach utilizes the well-known MIAS datasets to develop techniques for classifying breast cancer. The images within the MIAS collection are stored in portable gray map (PGM) format and possess dimensions of 1024×1024 pixels. The MIAS dataset comprises a total of 322 images that have been categorized into three distinct groups. Specifically, there are 61 images representing benign patients, 52 images representing malignant cases, and 209 images representing normal cases before DAA. The whole data is split randomly as 80-20 ratio for training and testing the details are shown in Table 2 and followed by the parameter setting in Table 3 [16, 21].

Table 2. Image details of the MIAS dataset before and after DAA

Table 3. Parameter setting

5.2 Performance evaluation metrics

Various classifiers can be assessed with factors like accuracy, precision, recall, and F1-score. Sensitivity and specificity are measures of categorization accuracy, which are determined by true positive (TP), true negative (TN), false negative (FN), and false positive (FP) outcomes. The confusion matrix in Figure 10 is an essential method in deep learning for assessing the effectiveness of image classification.

10.png

Figure 10. Confusion matrix

5.3 Confidence intervals

The formula used to determine the confidence interval based on a given metric is defined in Eq. (25).

$C I=x^{\prime} \pm Z \times \frac{\sigma}{\sqrt{n}}$  (25)

The k-fold cross-validation with 10 folds is recorded then the mean accuracy of the model 0.8990 for a level of confidence of 95% and the critical value Z=1.96.

The accuracy of classifying mass lesions in the dataset was improved with the use of freezing and fine-tuning approaches. The new model exhibited higher performance in comparison to four prior models in terms of accuracy, sensitivity, specificity, precision, area under the curve, and the F-score. Integrating the CNN via TL in the screening process leads to a significant improvement compared to previous methods. The findings indicated an accuracy of 96.12%, sensitivity of 91.5%, specificity of 94.3%, precision of 90.2%, F-score of 92.7%, and an AUC of 0.99. These findings surpass the other approaches listed.

The suggested approach may be applied in the future to diagnose other various cancer-related work. This work can also be implemented with real time patient detail to acquire more relevant accuracy of tumor prediction.