Residual block fully connected DCNN with categorical generalized focal dice loss and its application to Alzheimer’s disease severity detection

MRI dataset

The KAGGLE online community obtained the dataset (Dubey, 2023). The dataset obtained from KAGGLE consists of 5,121 axial images collected from different websites. The images in the dataset were collected and labeled into four different classes. These are no dementia, very mild dementia, mild dementia, and moderate dementia. There is no information about the age of the patients in the MR images obtained from the patients. The training dataset consists of MR images of 2,560 healthy (ND), 1,792 very mild dementia (vmD), 717 mild dementia (miD), and 52 moderate dementia (mD) individuals. The test dataset consists of MR images of 640 healthy (NC), 448 very mild dementia (vmD), 179 mild dementia (miD), and 12 moderate dementia (mD) individuals. The resolution of the images is 176 × 208. Samples are shown in Fig. 1. No other circumstances of the MR images obtained were specified.

Dataset pre-processing

Various image restoration techniques are used in the Kaggle AD dataset to increase the model’s training and test accuracy. First, the Non-Local Means (NLM) algorithm was applied to the images. Then, the image noise was de-noised using the weighted average of the pixel neighborhoods with similarity. Also, to improve performance when calculating the similarity of each pixel, instead of considering just one pixel, a small area of pixels is chosen around it, given by the small_window parameter. Patch_Size = 2 and Patch_Distance = 1 were selected to increase the noise removal performance of the NLM algorithm. Finally, Estimate_sigma is fused with the NLM algorithm to improve the performance of the proposed model.

The resulting images were passed through the CLAHE algorithm. Contrast-limited histogram equalization is performed by dividing images into small blocks called tiles in CLAHE in the OpenCV library. As a result of the experimental studies, the clipping level in CLAHE was chosen as 2.6. Samples of four different classes are shown in Fig. 2, showing the MR images in the Kaggle dataset before and after the noise removal and CLAHE. In Fig. 2, the pictures in the top line are the MR images before the image pre-processing techniques, and the images in the bottom line are the MR images obtained after the image pre-processing methods. The pre-processed training data set was doubled by applying random shifting, random rotation, random rescaling, and random horizontal and vertical flip methods.

The proposed categorical generalized focal dice loss function

The proposed Hybrid Loss function is a fused model of the Generalized Dice Loss (GDL) and Focal Loss (FL) (Sudre et al., 2017; Lin et al., 2020). The FL function formula is shown in Eq. (1).

As seen in Eq. (1), contrary to Cross-entropy (CE), Loss of Focus prevents suppression of CE Loss in large class imbalances. Thus, the negatives that make up most of the loss are reduced. Adding α to Focal Loss in Balanced Cross Entropy balances positive and negative samples. In addition, the adjustable focusing parameter γ ≥ 0 and the cross-entropy loss modulating factor (1 − p_t) was added to reduce the weight of easy-to-learn samples and to enable the deep-learning model to focus on difficult-to-train samples. As a result of the experimental studies performed in this study, we chose γ = 2 and α = 4, where p is the model’s predicted probability for the class (Lin et al., 2020).

The most significant disadvantage of Dice Loss is that although it has a good training score, its test score is low due to its poor response to class imbalances. As a solution to this, Sudre et al. (2017) transformed the Generalized Dice Score (GDS) function proposed by Crum, Camara & Hill (2006) to score multi-class segmentation into a loss function named Generalized Dice Loss (GDL). The GDL function formula is shown in Eq. (2). (2)${\displaystyle GDL=1-2\frac{\sum _{l=1}^2{w}_l\sum _n{r}_{ln}{p}_{ln}+\in }{\sum _{l=1}^2{w}_l\sum _n{r}_{ln}+p_{ln}+\in }}$

where r_l is the ground truth, and p_l is the predicted value. ∈ is a smoothing value for fine-tuning.

Here w_l isused to provide immutability to different labeled sets where w_l is formulated as shown in Eq. (3). (3)${\displaystyle w_l=\frac{1}{{\left(\sum _{n=1}{r}_{ln}\right)}^2}}$

where r_l is the ground truth. In this study, FL and GD, seen as robust models in multi-class segmentation, were fused to predict four different classes of AD, as shown in Eq. (4). (4)${\displaystyle Fusion\text{\_}Loss=\left(1-lamda\right)\ast GDL+lamda\ast FL}$

The lambda value was chosen as 0.5 because of experimental studies.

Performance metrics

Accuracy, Sensitivity, and Specificity metrics were used to measure the performance of the proposed deep learning-based classification model. Accuracy is the ratio of correctly predicted samples to the sum of all correct and incorrectly guessed samples; Sensitivity, or recall, is the ratio of positively correctly predicted samples to the sum of negative incorrectly predicted samples and correctly predicted positive samples. Specificity (Spec) represents the ratio of correctly predicted negative samples to the sum of correctly predicted negative and incorrectly predicted positive samples.

The proposed methodology

The end-to-end automatically AD classifier block diagram is shown in Fig. 3. The section including CLAHE covers the pre-processing stage of the dataset, and the section after CLAHE covers the training and testing of the model.

The proposed deep learning model

In the proposed deep learning architecture, 2xConvolutional2D blocks with 3 × 3 filters are used in all layers before the fully connected layer. In addition, the Xavier was used as the weight initializer function (Glorot & Bengio, 2010; Sharma et al., 2023; Nour, Senturk & Polat, 2023). Group Normalization (GN) is the normalizer, and ReLU is the activation function of the proposed model (Wu & He, 2020). Wu & He (2020) demonstrated with various experiments how superior the group normalization technique is to other normalization methods.

Besides, ADAM was used as the optimizer to optimize the weights of the proposed model (Ba & Kingma, 2015). In the last layer, Softmax was used as an activation function to find out which class the MR image belongs to the AD class. If we pay attention to the proposed model, it is seen that the U-shaped network is formed by taking the deconvolution part and adding the Fully connected network. Deep Neural networks are profound networks due to the multitude of layers. Therefore, the information learned in the first layers can be forgotten. The residual blocks (ResNet Block) ensure that the convolutional layers do not forget the information they learned in the first layers of the network. In the proposed model, residual blocks are followed by 2 × 2 maximum pooling (max-pool), in which the feature maps’ spatial dimensions (height and width) are reduced by half. Max-pool reduces the computational cost by reducing the number of trainable parameters. The layers of the proposed deep learning architecture are shown in Fig. 4. in detail.

Proposed model’s experimental studies

This section explains the experimental studies of the proposed deep learning architecture on the KAGGLE dataset in detail.

Employed hardware materials

Experimental studies of the proposed architecture were conducted with a computer equipped with Intel(R) Core(TM) i5-8300H @ 2.30 GHz CPU, 32 GB RAM, and NVIDIA GTX1050 4 GB GPU. In addition, the deep learning ecosystem consists of artificial intelligence libraries with Python 3.7 programming language based on Anaconda. Libraries and ecosystems are entirely open source. The dataset consists of training, validation, and test datasets. A total of 20% of the training dataset was used as a validation dataset using the scikit-learn library. We performed 5-fold cross-validation to measure the training dataset’s heterogeneity and the model’s fit.

Comparison of parameters

The model used in the proposed study and the models used in the other literature were analyzed comparatively regarding the number of parameters. As can be seen from the comparisons in Table 1, the proposed model is the one with the smallest parameter. It can be seen from Table 1 that the model with the smallest parameter is the recommended model.

Comparison of loss functions

The proposed fusion loss function has been analyzed in comparison with the most used mean absolute error (MAE), root mean squared error (RMSE), cross-entropy, FL, and GDL loss function for classification in the literature. As can be seen from Table 2, the proposed fusion loss function is the most suitable loss function for the proposed architecture.

Settings and studies on the KAGGLE dataset

Using our deep learning architecture, we performed binary classification, i.e., distinguishing NC and AD, and multi-class classification (NC, very mild AD, mild AD, moderate AD) experimental studies on the KAGGLE dataset. The deep learning network was trained for 15 thousand epochs for these two studies. After 15.000 epochs, the validation accuracy of the network remained constant at around 0.99. The minibatch size is 32, and ADAM is chosen as the optimizer. Numerical information about the data set used in training the proposed model is listed in Table 3. As can be seen from Table 3, there is a significant imbalance between classes. In this case, the benefits of data augmentation methods will be limited. Therefore, GDL and FL, which are successful and robust loss functions in class imbalances, are proposed as a solution to the class imbalance of the data set.

Experimental results and analysis

In this section, experimental results are shared, and comparative analyzes are made with other studies in the literature. In addition, information is given about the deep learning models used for detecting AD and how they train the dataset on the performance criteria they use. A slice-based data set was used in this study. The Kaggle dataset is a low-resolution and slice-based dataset collected on different websites. In addition, class imbalances are also very high in the dataset. Also, the Kaggle dataset’s having axial views is another challenge. For these reasons, the Kaggle dataset is very challenging. The most significant difficulty of all the literature and proposed research studies is distinguishing AD from NC in the MCI stage. If AD is successfully detected in the MCI stage, the effects of this neurodegenerative and irreversible disease can be slowed down. In addition, one of the most critical challenges is to detect NCs (very MCI) that may have AD. This study focused on two different experimental studies: (i) distinguishing NC from AD at the MCI stage; (ii) To successfully classifying MR images with NC, vMCI, MCI, and MC. For this reason, articles in the literature that made these two different experimental studies and analyzed their results were examined. Validation accuracy and validation loss values obtained using the categorical generalized focal membrane loss function of the proposed model are shown graphically in Figs. 5 and 6.

The comparative performance results of our proposed fusion loss-based deep learning architecture are shown in Tables 4 and 5. The proposed deep learning architecture achieved accuracy values of 0.973 in binary classification and 0.982 in multi-class classification. At the same time, it obtained high sensitivity and specificity values compared to other studies in the literature.