2.1

Deep Neural Network (DNN) application problems

The development of big data has driven the rise of DL, however, unlike natural image datasets, medical data is difficult to collect in a standardised way, and data annotation requires a high degree of specialisation and is expensive, making it very difficult to obtain large-scale annotated data. In addition, feature learning in deep networks is closely related to the class distribution of the dataset, and the very limited and unbalanced medical data is prone to problems such as biased feature selection and overfitting. Therefore, the application of DNNs to fine-grained chest X-ray aided diagnosis of high standard still faces two critical problems as follows ^7-8.

2.2

Thoracic medical imaging assisted diagnosis system

The thoracic medical imaging assisted diagnostic system in this paper consists of the following main areas.

2.3

Cross-Entropy loss function

Cross entropy is often used as a loss function in DL. The loss function reflects the distance between the output value y’ and the label value y. A smaller gap indicates a better fit.

In the logistic regression task, let the input be x and the output be y’, the linear regression model can be abbreviated as

and define a non-linear mapping function from the input space to the output space as in equation (2).

For a 2 classification problem, where the label value y takes the value 0 or 1, the posterior probability function at y = 1 can first be defined as

The posterior probability at y = 0 is then defined as

In summary, the posterior probability P(y|x) can be defined as follows.

According to the maximum likelihood estimate, if all samples satisfy independent identical distribution, then a set of parameters can be identified to maximise the value of equation (5). Since the logarithmic function is monotonically increasing, then maximising the value of P(y|x) can be equated to maximising the value of log(P(y|x)), so taking the logarithm of equation (5) yields

The objective of logistic regression is to maximise this function, which adequately reflects the model performance error. The loss function can therefore be made to be the opposite of the above function, and for m samples the loss function is the cross-entropy function.

3.1

Experimental environment and parameter settings

The PyTorch framework is a commonly used DL framework in which one can design one’s own network model and automatically solve the gradient by back propagation within the framework. The reduction of the U-Net network framework and the training of the DB-U-Net network are based on the PyTorch framework. A pre-configured runtime environment and library files are used to run the lung CT image pre-processing algorithm and the graph cut algorithm.

The basic configuration of the experimental environment in this paper is an Inteli5-8400 processor, two NVIDIAGeForce1070Ti graphics cards (8GB of video memory), 32G of RAM, and Ubuntu 16.04 Server Edition. The GPU parameters, such as computational power and speed, will affect the network training of the experiments, and the use of GPU performance will reduce the training time of the split network in this paper.

The experimental parameters are shown in Table 1. Because the lung CT images used in this paper are relatively large, the number of samples selected for the same training session will have the problem of overflowing video memory, which will affect the calculation results and the calculation speed. The effectiveness of the training depends heavily on the learning rate, which is set to 0.01, 0.001, 0.0001 and 0.00001, respectively.

Table 1.

Experimental parameter settings.

Stochastic gradient descent is used for training the model in this chapter to solve for the minima, updating the model weights in each iteration until it gradually converges to the minima. In a single iteration, the weights of the model are updated using a batch size number of samples each time. The advantages of using stochastic gradient descent over other optimisation algorithms are as follows: (1) it is possible to train the network model without loading all the data into memory at once, taking up less memory and video memory; (2) stochastic gradient descent algorithms train the network faster; (3) within reason, increasing the batch size can make the direction of gradient descent more accurate and faster, as smaller batches can cause the model to to fall into other locally optimal solutions. When training the NN, the batch size was set to 16.

3.2

General flow of the experiment

In this paper, a feature interpretability-guided multi-scale integrated convolutional NN is proposed to achieve abstract feature reuse with understanding of cnn depth features, and to accurately and automatically perform the task of classifying true and false recurrence of glioma in DTI images. Firstly, three single classification models for true and false recurrence of glioma were trained separately using the classical cnn framework as the base model. All feature maps are then visualised layer by layer, and the heatmap mapping approach is used to visually capture the respective “focus” of the different layers of the single classification models. For example, some layers focus more on edge information, while others highlight differences between pixels, etc. Based on this, the three monoclassification models are empirically observed by the imaging practitioner to locate the most relevant layer for each glioma lesion area, visually completing the feature selection process. Finally, the selected multi-scale features from different networks were fused to construct an integrated model of true and false recurrence of glioma. The experimental flow chart is shown in Figure 1.

Figure 1.

Experimental flow chart.

3.3

Performance evaluation indicators

In order to verify the accuracy of the algorithm based on DL and feature mixing, the performance of the algorithm was measured based on the categories to which the sample labels belonged compared with the actual predicted categories of the algorithm using the evaluation criteria shown in Table 2 to overcome the singularity and one-sidedness of the algorithm performance evaluation in the field of medical aid diagnosis in terms of accuracy alone.

Table 2.

Evaluation criteria for detection of lung nodules.

The meanings of the indicators in Table 2 are as follows.

In the lung nodule detection algorithm, the ultimate aim is to extract all the lung nodules in the lung CT image sequence, but since interference from other tissues within the lung often causes a degree of false positives, the experiment uses the accuracy (ACC), sensitivity, and specificity (Spe) defined by the four criteria mentioned above for a comprehensive measure of the overall discriminatory ability of the algorithm.

The accuracy is used to measure the overall classification capability of the algorithm, and the higher the accuracy, the better the overall classification of the algorithm, as defined in equation (8).

Sensitivity, also known as true positive rate and recall, is used to measure the ability of the algorithm to discriminate between regions of lung nodules, i.e. the ratio of the number of correctly predicted samples among all nodule samples, with higher sensitivity representing a lower rate of missed detections by the algorithm, as defined by the formula in equation (9).

Specificity is used to measure the ability of the algorithm to discriminate between non-nodular regions, i.e. the ratio of the number of correctly predicted samples out of all non-nodular samples, with a higher specificity representing a better determination of non-nodular regions, as defined by the formula in equation (10).

In this paper, we use the FROC (Free Response Receiver Operating Characteristics) criterion to measure the detection effectiveness of the algorithm by calculating the cpm (Competitive Performance Index) value, where cpm is the average number of false positives per set of CT images (FPs per scan) i.e. the average detection rate at horizontal coordinates of 1/8, 1/4, 1/2, 1, 2, 4, 8. The following experiments use cpm values as an assessment criterion.

4.1

Experimental data sources

The dataset was derived from lung CT image segmentation, with 7,500 CT images, 6,000 for training and validation, which contains 4,800 training set data and 1,200 validation set data. The training and validation datasets were all collected from the National Centre for Biological Information and contained a dataset of 2000 normal CT line images, 2000 CT line images of viral pneumonia and 2000 CT line images of positive new coronary pneumonia as experimental subjects. Details of the dataset are shown in Table 3 for a total of 6000 CT images of lung data, 80% of the data is the training set data and 20% of the data is the validation set data.

Table 3.

Lung CT image dataset

4.2

Analysis of experimental results

To validate the effectiveness of the proposed DL and feature mixing based lung nodule detection model, CT image sequences were sent to the network for detection, and patient image data from the test set were selected for the following self-comparison experiments and statistical quantification of the prediction results by calculating cpm values using guidelines to measure the performance of the algorithm: (1) Lung nodules were detected using the traditional R-fcn network and the results were recorded as modell. Table 4 shows the difference in performance between the three network structures on the LUNA16 dataset for Figure 2 shows the FROC curves of the three models for the improved process.

Figure 2.

Comparative analysis of the detection performance of the algorithm improvement process

Table 4.

Algorithm improvement process detection performance comparison

The process of the change in the average detection rate of the algorithm during the improvement can be seen by observing Table 4 with Figure 2. By upgrading the RcsNet to denseNet with fpn for feature extraction, the algorithm detection sensitivity was improved and the recognition performance of the network was enhanced. Then according to the results of K-meaning is clustering change the settings in the rpn network with and correlation in the network, and introduce the feature pyramid structure, training using Focalose loss function instead of the traditional cross-entropy cost function to further improve the recognition performance of the network, have a higher detection accuracy for small-scale nodules, so that the network has better results. The above experiments show that the detection sensitivity of the algorithm is improving in the process of improvement, and the average detection rate can reach 96.6% when the average number of false positives in each group of CT images is 8. The recognition performance of the proposed algorithm is better than the original method in the detection of lung nodules.

In summary, the following conclusions can be drawn from the above experiments: even though lung nodules have a series of complex characteristics such as varying sizes, complex shapes and uncertain locations, the proposed DL and feature mixing based lung nodule detection algorithm still shows good detection results for nodules of different scales and good discrimination ability in the face of complex lesion areas. The network not only has a high prediction probability for larger diameter lung nodules to reach a correct judgment, but also for smaller diameter microscopic lung nodules to reach a correct prediction conclusion.

Table 5 shows the performance of the algorithm in this paper compared with the FasterRCNN, the multi-view convolutional NN model and the multi-view deep belief network for lung nodule detection. The FROC curves for the four models are shown in Figure 3.

Figure 3.

Comparative analysis of lung nodule detection performance by model

Table 5.

Comparison of lung nodule detection performance by model

Analysis of Table 5 and Figure 3 shows that the proposed algorithm performs well in the task of lung nodule detection, and its overall sensitivity is better than other algorithms, with a sensitivity of 96.6% at an average of 8 false positives per group of CT images. Through the above analysis, the improved algorithm outperforms the traditional algorithm in terms of feature extraction ability and discrimination ability, which can effectively lay the foundation for the subsequent diagnosis of the imaging physician and achieve the purpose of assisting the physician in the diagnosis.