3.3. Our Proposed Custom ResNet-14 Architecture

In this section we will briefly explain the proposed CNN custom Residual ResNet architecture. After all the pre-processing steps above the authors have built an end-to-end model by modifying the original version-I residual ResNet-18 [31], into proposed ResNet-14 network structure as it illustrated in Figure 5. The MR image with dimension 75 × 75 × 1 is provided as input layer in the structure. We added batch normalization (BN) [49] in the model before the activation function rectified linear unit (Relu) and right after convolutional layers (Conv) with 3 × 3, which acts like a regularization. The vanishing gradient problem is reduced significantly through this operation. In addition to this, a sequence of 3 inner ResNet stacks of convolutional with stride 2 of max pooling 3 × 3 with n = 2 parameters instead of 3 to avoid the overfitting. There are totally 6n + 2 stacked weighted layers.

Our customized ResNet-14 architecture.

Further, we are used to controlling the learning process with fine-tuned hyper-parameters by manually having a great impact on the performance of the model. In the complied stage on the proposed architecture, we have chosen the Adam [50] optimizer, which can keep tracks of an exponentially decay average. The learning rate was configured to be set dynamically on the basic of the number of epochs, batch size to 32 and the learning rate is 0.001 as in our case we used with 120 epochs. At the ends, 3 fully connected layers (FC) with average pooling (Avg pool) and softmax activation function have been added to detect the healthy, partial and rupture tears in the MRI. The details of the convolutional layers and their order in the custom ResNet-14 model in the Table 3. The total number of parameters are 179,075.

The configuration detail of customized ResNet model-14 with their output size.

Finally we involved the real-time data augmentation in our model, which generated different images after running each epoch. It randomly augmented the image at runtime and applied transformation in mini-batches [51]. So, it is more efficient than offline augmentation because it does not require extensive training. The technique of offline data augmentation significantly increased the diversity of their available data without actually collecting new data by cropping, padding, flipping, rotating and combining in the case of Alzheimer’s stage detection, brain tumor and others in the MRI [52,53,54].

The real-time data augmentation performed good accuracy with the CNN inception v3 model for breast cancer [55]. We used real time data augmentation with a class Image_Data_generator which generated batches of tensor image data [56,57,58] from the keras library. The following Table 4, describes about augmentation parameters which we used in the real time augmentation.

List of selected real-time augmentation with arguments and their description.

Furthermore, the block diagram of the proposed work’s whole process is illustrated in Figure 6, with four main stages. Firstly, the data input stage, where the image dimension is combined with metadata to generate images through the pickle library. In the second stage, the images are resized through the region of interest and then applied with hybrid-class balancing. The model building stage is done through our custom ResNet-14 with and without online data augmentation. In the last stage, the performance is measured and compared through random train/test split and K-fold cross-validation to detect anterior cruciate ligament tear.

A block diagram of the proposed methodology.