Lung cancer is a deadly and malignant condition. There are numerous models that can identify lung cancer, but their accuracy must be improved. Therefore, it's critical to identify lung cancer with great precision and sensitivity. A new model called Maximum Sensitivity Neural Network (MSNN) has been offered as a solution to this issue. The suggested model beat existing deep learning models with 96.9% accuracy and 94.6% sensitivity after experimental verification. Sensitivity map has been superimposed on lung ct scan image to show the area of an image that contributes the most for classification of nodule as cancerous or noncancerous. As a result, this innovative approach has good performance in identifying lung cancer with a low false positive rate, which can boost the accuracy of medical diagnosis.
Research Article
Enhancing Early Detection of Lung Cancer: Image Processing and Nodule Classification using Sensitivity maps
https://doi.org/10.21203/rs.3.rs-3397247/v1
This work is licensed under a CC BY 4.0 License
Version 1
posted
You are reading this latest preprint version
CT Scan Image
Deep Features
Lung Cancer
Maximum Sensitivity
Sensitivity Maps
Lung cancer can be identified using a variety of methods, including CT (Computed Tomography) scans, MRI (Magnetic Resonance Imaging) scans, and others. CT scan images are preferable above all other modalities because they can detect a very small nodule in the lungs. Patients can receive early treatment if the disease is detected early, minimizing the number of deaths.
In this work, the preprocessing stage helps in noise reduction from images wherein, the Median Filter beats the Average, Gaussian, Laplacian, and Wiener Filters. Following that, nodule extraction and classification were performed using several image segmentation methods such as K-Means Clustering, Watershed Segmentation, and Otsu Thresholding. Various performance parameters were calculated to validate the proposed model results, and it was observed that the model has the best performance. At last, geometric features have been calculated for the detected nodules which also validate the results of software. This research contributes to the advancement of lung cancer diagnosis and early intervention.
When dealing with a range of diseases, accurate diagnosis becomes a difficult task in medical science. Although various forms of cancer have been identified, lung cancer is the most dangerous and life threatening. According to a report published by National Cancer Registry Programme, India, approximately 7,84,821 deaths/year happen due to cancer [1]. When the human body fails to produce healthy cells, a tumour/nodule forms which is also known as neoplasma. It can be benign or malignant. Early detection of lung nodule or pulmonary nodule can considerably improve the effectiveness of treatment and hence raises the survival rate of patients.
CT scan is the most effective imaging modality for finding lung cancer. Internal organs, soft tissue, bones, and blood vessels are all captured in fine detail. It also offers a high resolution, low distortion, and high contrast, allowing for more detailed analysis of the lungs and faster detection of lung nodules [2].However, CT scan images are prone to noise, which affects image quality and, as a result, reduces radiologist efficiency in detecting lung cancer at an early stage. Therefore, deep neural networks have made significant advancements in the field of intelligent medicine in recent years.
The following are the main contribution of this paper:
-
A deep convolution neural network called MSNN has been proposed which finds and learns patterns in lung CT scan images to detect lung cancer. The proposed architecture has a small number of parameters and can be utilized for limited training dataset. The model also used a global average pooling layer (GAP) to overcome the problem of overfitting.
-
Another important feature of this model is the use of sensitivity maps which are plotted to visualize which part of the image contributes more to the classification of the image.
-
The effectiveness of the proposed model for deep feature extraction has been investigated. The model can extract deep features of the image which later is also fed into classifiers like the K-Nearest Neighbor (KNN) and Support Vector Machine (SVM) for classification and comparison is made between them.
-
To obtain good performance of different classifier, bayesian optimization method has been used to select the optimal parameters for the different classifiers. As a result, the proposed model has a high accuracy for lung cancer detection.
It is essential to detect a lung nodule with high accuracy. This is a difficult task for radiologists and more enquiry time is required. Therefore, a variety of deep neural networks have been developed for lung cancer detection and proved to be better than artificial neural networks [3–5]. Some of them are studied and summarized below:
-
Zhang et al. [6] used deep neural network to extract shallow and deep features for ovarian cancer classification.
-
Wu et al. [7] designed convolutional neural network (CNN) which is based on AlexNet architecture for categorization of ovarian cancer pathological images, and the model accuracy rate was 78.2%.
-
Tajbakhsh and Suzuki [8] applied artificial neural networks and convolutional neural networks for testing benign and malignant nodules in lung CT scan images. The experimental results showed that when it came to categorising lung lesions and tumours, convolutional neural networks outperformed other artificial neural networks.
-
Wei Shen et al. [9] designed a convolutional network named Multi Crop CNN for classification of lung nodule malignancy by using learned deep features.
-
A. Masood et al. [10] proposed DFCNet model which is a convolution neural network used for classification of lung nodule into four stages. This proposed network gave an accuracy of 84.58% indicating the effectiveness of proposed method.
-
Satyavratan Govindarajan and Ramakrishnan Swaminathan [11] proposed a CNN model with optimized parameters to detect COVID-19 with sensitivity of 97.63% and F-measure of 97.1%. The model proved to be efficient in providing visual diagnostic solution.
-
Long et al. [12] developed new model which is a convolutional neural network. In this model, fully convolutional layer is used in place of fully connected layer. This modification enabled the precise pixel-by-pixel prediction of the entire image in a single forward pass.
-
In [13], Christ et al. segmented the liver by cascading two fully connected networks (FCN), with the first FCN segmenting the liver as the region of interest for the second FCN segmenting the lesions in the liver. This method successfully segmented lesions in CT images with a Dice score of 0.823 and MRI images with a Dice score of 0.85.
-
To limit the number of false positives brought on by the imbalanced ratio of background and foreground pixels in medical images, Zhou et al. [14] proposed to apply the focal loss on the FCN. In this structure, the intermediate segmentation results were generated using the FCN, and the false positives were subsequently eliminated using the focal FCN.
3.1. MSNN architecture
A private database containing lung CT scan images is used in this work. The goal is to create a novel, effective convolution neural network architecture that detects lung cancer and outputs a binary data to detect cancerous nodule or noncancerous nodule. Each input CT scan image is a member of a specific class, and a probability score is assigned to each output image.
MSNN has been designed for the diagnosis of lung cancer by using lung CT scan images. To classify the test images, it learns patterns from lung CT scan images. To identify patterns in the images, this network input layer accepts grayscale images of size 512X512. Figure 1 presents MSNN architectural layout which shows that it is made up of five successive blocks. Block1-block4 is made up of four layers namely convolution(conv), Batch Normalization (BN), Rectified Linear Unit (ReLU) and max pooling layer. Wherein, block5 is made up of convolution(conv), Batch Normalization (BN), Rectified Linear Unit (ReLU) and global average pooling layer. Further, table I also displays the MSNN architecture details with total parameters for each layer.
The first layer is convolution layer, which performs convolution operation between the input image(f) and filter size (g) by using Eq. 1:
Where x and k are spatial variables.
In general, a smaller filter size may lead to an overfitting issue, while a bigger filter size may increase the underfitting issue. Therefore, this layer uses 8 filters with a 6x6 ideal filter size.
The next successive layer is BN layer which expedites training speed and lessens network sensitivity. Therefore, performing normalization over a batch(v) of m instances for ‘i’ unit can be done using the following steps.
Firstly, compute batch mean by using Eq. (2):
Where r ranges from 1 to m
Secondly. Compute batch variance by using Eq. (3):
Thirdly compute normalized batch instances by using Eq. (4):
Lastly, scale with learnable parameters by using Eq. (5):
ReLU layer helps to add nonlinearity to the network by adding a rectifier function which is computing linear operations during convolution. The function works by using Eq. (6):
Max pooling layers help to decrease the size of the convolved feature map to reduce computational costs.
Global Average Pooling layer is utilized in this work which helps to enhance the model performance by reducing the loss function value.
Each input image is given a probability score using a series of FC (Fully Connected), BN, ReLU, FC, and SM (Soft Max) layers. Where FC layer helps in the classification of images into categories. And SM layer converts the output of the last layer of network into a probability distribution.
Several experiments have been carried out using lung CT image datasets to demonstrate the effectiveness of MSNN, and comparison with other models have been done.
4.1. Dataset and training options
Since classification of test images depends on several parameters, neural network training is quite crucial and needs to be carefully considered. The following setting has been done while training MSNN. Adam optimizer with optimum batch size has been chosen so that memory space can be saved, and elapsed time is not too high. Similarly, the network will experience overfitting if the epoch value is too high. Conversely, a low epoch value will cause the network to converge quickly and cause training to end early. Therefore, batch size of 20 and epoch value of 20 has been set to train MSNN. The dataset is divided into training and validation set randomly for split1. 80% of images from datasets are used for training and the remaining images are used for testing. Split2, Split3, and Split4 imply that 70%, 75%, and 85% of the database images are used for training respectively, while the remaining images are used for testing.
In this work, lung CT scan images in DICOM format have been acquired from A.J Hospital and Research centre to test the performance of MSNN. The dataset consists of 3252 lung CT scan images from 145 patients, in which 1652 lung CT scan images belong to 72 patients with cancer and the remaining 1600 lung CT scan images belong to 73 patients with healthy lungs. Sample images from the dataset are displayed in Fig. 2.
Table I: Details of MSNN architecture
| Type of Layer | Output layer size | Number of Filters used | Size of Filter | Stride Value | Padding | Number of Parameters |
|---|---|---|---|---|---|---|
| Convolution | 508X508X8 | 8 | 6X6 | 1 | [1 1 1 1] | 296 |
| BN | 508X508X8 | - | - | - | - | 16 |
| ReLU | 508X508X8 | - | - | - | - | 0 |
| Max pooling | 254X254X8 | - | - | 2 | [1 1 1 1] | 0 |
| Convolution | 254X254X16 | 16 | 4X4 | 1 | [1 1 1 1] | 2064 |
| BN | 254X254X16 | - | - | - | - | 32 |
| ReLU | 254X254X16 | - | - | - | - | 0 |
| Max pooling | 127X127X16 | - | - | 2 | [1 1 1 1] | 0 |
| Convolution | 127X127X32 | 32 | 4X4 | 1 | [1 1 1 1] | 16864 |
| BN | 127X127X32 | - | - | - | - | 64 |
| ReLU | 127X127X32 | - | - | - | - | 0 |
| Max pooling | 63X63X32 | - | - | 2 | [1 1 1 1] | 0 |
| Convolution | 63x63x64 | 64 | 4X4 | 1 | [1 1 1 1] | 32832 |
| BN | 63x63x64 | - | - | - | - | 128 |
| ReLU | 63x63x64 | - | - | - | - | 0 |
| Max pooling | 31X31X64 | - | - | 2 | [1 1 1 1] | 0 |
| Convolution | 31X31X128 | 128 | 4X4 | 1 | [1 1 1 1] | 131200 |
| BN | 31X31X128 | - | - | - | - | 256 |
| ReLU | 31X31X128 | - | - | - | - | 0 |
| Max pooling | 15X15X128 | - | - | 2 | [1 1 1 1] | 0 |
| Convolution | 15X15X256 | 256 | 4X4 | 1 | [1 1 1 1] | 524544 |
| BN | 15X15X256 | - | - | - | - | 512 |
| ReLU | 15X15X256 | - | - | - | - | 0 |
| Max pooling | 7X7X256 | - | - | 2 | [1 1 1 1] | 0 |
| Global Average pooling | 1x1x256 | - | - | - | - | 0 |
| Fully Connected | 1x1x512 | - | - | - | - | 131584 |
| BN | 1x1x512 | - | - | - | - | 1024 |
| ReLU | 1x1x512 | - | - | - | - | 0 |
| Fully Connected | 1x1x2 | - | - | - | - | 1026 |
| SoftMax | 1x1x2 | - | - | - | - | 0 |
4.2. Performance metrics
The performance of MSNN has been evaluated by obtaining confusion matrix which helps in calculating all the metrics related to performance of network.
-
Accuracy
This parameter defines true positive and true negative or correct cases over total number of cases which includes false cases too. Eq. (7) below can be used for accuracy calculation:
-
Precision
This parameter determines true positive cases over true positive and true negative cases. Eq. (8) below can be used for precision calculation:
-
Recall/Sensitivity
This parameter measures true positive rate. A value close to 100% means its test result is positive and patient has disease. Eq. (9) below can be used for recall calculation:
-
F-Score
This parameter determines how many cases are classified correctly. Eq. (10) below can be used for calculating this parameter:
-
Specificity
This parameter determines how many cases are correctly classified as negative. Eq. (11) below can be used for calculating this parameter:
To evaluate the performance of MSNN completely, a private dataset has been acquired from the hospital and further performance metric parameters have been calculated for four different splits which are shown in Table II. The confusion matrix has been plotted in Fig. 5 for four different splits of dataset.
4.3. Superimposing Sensitivity map on lung CT scan image
In this work, sensitivity maps [15] have been superimposed on lung CT scan images to find out which part of CT scan image are important and helps in taking right decision for classification. Sensitiivty map highlights the regions in red, which represent the most significant area in the model's decision-making process. While regions in blue or other colours indicates that they are less important to the model's conclusion. This work uses binary classification, meaning that each detected nodule only falls into one of two categories: malignant or not. A probability score of 0 suggests the model is very confident in determining that the image does not belong to a particular class, whereas a score of 1 indicates the model is very confident that the instance belongs to the positive class. Figure 3 shows that it is a malignant image since the probability is 1 for the cancerous class and 0 for the non-cancerous class.
4.4. Deep feature extraction using MSNN model
In this research work, the proposed model extracted deep layer features. These features contain a lot of distinguished information about the image. Therefore, by using lung CT scan image, features of different convolution layers have been extracted which is shown in Fig. 7. Convolution layers can extract features from various angles and therefore, the first convolution layer extracts fundamental information which contains spot and edges. Results showed that deep layers extracted high level and abstracted features by merging earlier features.
Hence, features retrieved from the deeper layers are more suited for classification[16]. The extracted features are then fed as an input to KNN (K-Nearest Neighbor) and SVM classifier. While using classifier different parameters need to be set manually, but this approach takes a lot of time. So, bayesian optimization approach has been utilized to select optimum parameters automatically. It is a process that optimizes an unknown objective function by iteratively evaluating the objective function at a sequence of points.
Figure 4. shows the graph with loss function (minimum objective) plotted on Y-axis and function evaluation plotted on X-axis. The loss function determines the difference between the predicted values and the true values of a model. Wherein, function evaluation refers to the number of times the algorithm is applied during the validation process. In this work, Mean Square Error is used as a loss function. The loss function is clearly minimized in (c) after applying the MSNN classifier a few times.
After establishing the best parameters through Bayesian optimisation and training the various classifiers. The performance of the classifiers on the test set is assessed. The effectiveness of the classifiers is assessed using a few performance indicators, including sensitivity, specificity, precision, F1-score, and accuracy.
Table II: MSNN Performance for different splits of dataset.
Accuracy and loss plots have been shown in Fig. 6 respectively for split1 only. Performance comparison of Proposed MSNN with other existing deep learning models has been shown in Table III.
Table III: Performance comparison of MSNN with other existing deep learning models
Generally, medical data will be so large that overfitting is a constant issue for deep neural networks. To overcome this problem, a global average pooling layer is used which reduces network complexity.
The proposed MSNN model is a deep convolution neural network which is effective in identifying lung cancer. The model is trained with a small dataset consisting of lung cancer CT scan images obtained from hospital. A small dataset can cause model overfitting as it may not contain sufficient variations and complexities which are required to make the model learn different patterns. As a result, models can be inefficient to classify the test data correctly. Therefore, Global Average Pooling layer are utilized in this work which helps to enhance the model performance by reducing the loss function value [17]. Reduced loss function signifies that model can do predictions with more accuracy.
As a result, model performance has been improved with accuracy of 96.9% and sensitivity of 94.6%. Hence, it can be defended that MSNN is most suited for lung cancer detection because the accuracy stated in previous study [18] is 88%, respectively.
4.5 Calculating geometrical parameters for detected nodule.
According to the radiologist, nodules might be oval, lobulated, polygonal or ragged in shape [19]. Nodules with round or oval margins are referred as ‘round'; distinct margins with some smooth and large convexities are referred as ‘lobulated', while a margin in polygonal shape is referred as ‘polygonal. With irregular margin these are referred as ragged shape.
Figure 8 shows that different geometrical parameters has been calculated for deteced lung nodules to support radiologist decision.
High accuracy is essential while building a deep learning model for detecting lung cancer when medical applications are considered. Therefore, in this paper, MSNN model has been designed carefully which exhibits a respectable level of accuracy of 96.9% and sensitivity of 94.6%. This is possible if the model complexity can be lowered, which is accomplished by adding a global average pooling layer in the last stage. To further examine the performance of model, extracted features has been fed to classifers like KNN and SVM for classification. To enhance the performance of classifiers, Bayesian optimization technique has been utilized to select the optimal hyperparameters for different classifiers. After analysis, it is observed that MSNN classifier outperformed other classifiers in this work since loss function was minimum compared to other models. Performance of model has been analyzed by calculating sensitivity, specificity, precision, accuracy and F1 score parameters. A sensitivity map has been plotted for input lung CT scan image that accurately identifies the nodule area in red colour. This further aids in the classification of the image into cancerous and non-cancerous images. Atlast, few geometrical parameters such as area occupied by nodule,its perimeter and shape has been calculated to support radiologist decision.
ACKNOWLEDGEMENTS
Authors would like to thank Dr.Ashwin Polnaya, Department of Radio-Diagnosis, Interventional radiology division, A.J. Institute of Medical Sciences, Mangalore (India) which helped in acquiring database of existing patients from hospital and guided to understand raw medical images.
- R. L. Siegel, K. D. Miller, H. E. Fuchs, and A. Jemal, “Cancer Statistics, 2021”, Cancer Journal for Clinicians, vol. 71, no. 1, pp 7–33, 2021.
- Prashant Mathur, Krishnan Sathish Kumar, “Cancer Statistics, 2020: Report from National Cancer Registry Programme, India”, National Centre for Disease Informatics and Research, Bengaluru, India, vol. 6, no. 6, pp 1063–1075,2020.
- M. Savic, Y. Ma, G. Ramponi, W. Du, and Y. Peng, “Lung Nodule Segmentation with a Region-Based Fast Marching Method”, Journal of sensors, vol. 21, no. 5, pp 1–32,2021.
- Mohd Firdaus Abdullah, Muhammad Safwan Mansor and Siti Noraini Sulaiman, “A Comparative Study of Image Segmentation Technique applied for Lung Cancer Detection”,IEEE International Conference on Control System, Computing and Engineering, Penang, Malaysia, pp 72–77, 2019.
- M. Vas and A. Dessai, “Lung cancer detection system using lung CT image processing”, International Conference on Computing, Communication, Control and Automation, ICCUBEA,2018.
- K. Senthil Kumar, K. Venkatalakshmi, and K. Karthikeyan, “Lung Cancer Detection Using Image Segmentation by means of Various Evolutionary Algorithms”, Computational and Mathematical Methods in Medicine, pp 107-114, 2019.
- P. M. Selvi and B. Ashadevi, “Elimination of Noise in CT Images of Lung Cancer using Image Preprocessing Filtering Techniques”, International Journal of Advanced Science and Technology, vol. 29, no. 4s, pp 1823–1832, 2020.
- S. Makaju, P. W. C. Prasad, A. Alsadoon, A. K. Singh, and A. Elchouemi, “Lung Cancer Detection using CT Scan Images”, International Conference on Smart Computing and Communications, Kurukshetra, India, vol. 125, pp 107–114, 2018.
- A. I. Khan, J. L. Shah, and M. M. Bhat, “CoroNet: A deep neural network for detection and diagnosis of COVID-19 from chest x-ray images”, Computer Methods Programs in Biomedicine, vol. 196, 2020.
- Y. Jalali, M. Fateh, M. Rezvani, V. Abolghasemi, and M. H. Anisi, “ResBCDU-Net: A Deep Learning Framework for Lung CT Image Segmentation”, Journal of Knowledge Based Systems, vol. 21, no. 1, pp 1–24,2021.
- X. Huang, Q. Lei, T. Xie, Y. Zhang, Z. Hu, and Q. Zhou, “Deep Transfer Convolutional Neural Network and Extreme Learning Machine for lung nodule diagnosis on CT images”, Journal of Knowledge-Based Systems, vol. 204, pp 8357-8361,2020.
- A. Asuntha and A. Srinivasan, “Deep learning for lung Cancer detection and classification”, Multimedia Tools and Applications 2020 79:11, vol. 79, no. 11, pp 7731–7762, 2020.
- M. S. Rahman, P. C. Shill, and Z. Homayra, “A New Method for Lung Nodule Detection Using Deep Neural Networks for CT Images”, International Conference on Electrical, Computer and Communication Engineering, 2019.
- L. M. Pehrson, M. B. Nielsen, and C. A. Lauridsen, “Automatic Pulmonary Nodule Detection Applying Deep Learning or Machine Learning Algorithms to the LIDC-IDRI Database: A Systematic Review,” Diagnostics (Basel), vol. 9, no. 1, Mar. 2019.
- S. Saraswathi and L. M. I. Sheela, “Detection of Juxtapleural Nodules in Lung Cancer Cases Using an Optimal Critical Point Selection Algorithm,” Asian Pacific Journal of Cancer Prevention: APJCP, vol. 18, no. 11, pp 3143–3148, 2017.
- N. S. Nadkarni and S. Borkar, “Detection of lung cancer in CT images using image processing,” Proceedings of the International Conference on Trends in Electronics and Informatics, pp 863–866, 2019.
- S. Ren, K. He, R. Girshick, and J. Sun, “Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks,” IEEE Trans Pattern Anal Mach Intell S. Ren, K. He, R. Girshick, and J. Sun, “Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks,” IEEE Transaction Pattern Analysis and Machine Intelligence, vol. 39, no. 6, pp 1137–1149, 2017.
- X. Wang, L. Wang, and P. Zheng, “SC-Dynamic R-CNN: A Self-Calibrated Dynamic R-CNN Model for Lung Cancer Lesion Detection,” Computer Mathematical Methods Medicine, 2022.
No competing interests reported.