Visual Explainability-Driven DL framework for Lung Nodule Classification

Visual Explainability-Driven DL framework for Lung Nodule Classication

I.Tejaswini, T.Thanmai, M.Apphia, M.Mohit, and A.Sagar

Anil Neerukonda Institute Of Technology and Sciences

Abstract. The large number of images and the subtle characteristics of pulmonary nodules make it challenging to detect the lung cancer from CT scans. Complex and small nodules may provide a less accurate diag- nosis, and manual examination is time-consuming and result to variation among observers. Previous studies have shown excellent accuracy but of- ten fall short in terms of organization and understandability. This paper introduces an automated and explainable deep learning framework for lung nodule classication. CNN is used to classify CT scans. We use con- sistent preprocessing methods to make sure that model is strong and de- livers reliable results. Experiments conducted on public lung CT datasets show an accuracy of about 95%.

Keywords: Deep Learning · CNN · Machine Learning · Nodule Classi- cation

1. Introduction

   Lung cancer continues to be signicant health issue worldwide, leading to a signicant number of deaths related to cancer annually. Detecting pulmonary nodules early is crucial for improving survival rates; however, analysing lung CT scans is challenging due to the vast amount of imaging data and the subtle char- acteristics of nodules. These nodules capable to vary in size, shape, texture, and location. This variation makes manual diagnosis which subject to human error and time-consuming. As a result, there is an increasing need for automated and dependable computer-aided diagnosis (CAD) systems that can help radiologists accurately and consistently detect and classify lung nodules.

   Convolutional neural networks [3] (CNNs) have signicantly improved the medical image processing in our model by impulsively classifying lung nodules from the computed tomography [2] (CT) scans using contemporary deep learning architectures to keep the focus on the anatomically relevant body region, this work proposes a holistic deep learning pipeline for the automated segmentation of the lungs the standardization of image preprocessing.

   Through the usage of saliency maps and other visual explanation tools, it is possible to highlight the features that guide the CNN in making predictions. The results can be assessed by clinicians, and diagnostic reports may be created through the upload of the images and the usage of a web-based GUI. The usage of public lung CT image datasets has recorded an accuracy of about 95%.

   In the past few years, deep learning has recorded outstanding advancements in the area of computer vision, automatically identifying the most important features in images. This is contrast with the traditional approaches that relied on the manual specication of the features. CNNs have the capability of learning spatial and texture features in images. However, the CNN is a black box, and there is a need to increase the importance of the development of explainable deep learning models with ability to achieve the high accuracy and visual insights.

2. Literature Survey

   Detection of lung nodules is one of the prominent issues, especially with the in- crease in lung cancer cases, emphasising the importance of early detection using CT scans. Initially, conventional machine learning models such as Support Vec- tor Machine[6] (SVM), Random Forest[10], and Naive Bayes[4] were employed, which required hand-crafted feature extraction models such as texture, shape, and intensity features.

   Liu et al. [1] delivered a real-time deep learning- based detector called as STBi-YOLO, which is specically designed to detect the lung nodules using CT scans. This method is based on the YOLO framework, where stochastic pooling is incorporated, enabling a feature pyramid to eectively utilize the multi-scale information. Evaluation is done using the dataset containing annotated CT scan images, where performance is measured using detection real-time inference speed and accuracy. This method is initiate to have improved performance compared to other YOLO variants, where focus is only on detecting nodules, not classifying them.

   Nguyen et al. [2] proposed a framework using Faster R-CNN, where Adap- tive Anchor Boxes were incorporated to eectively detect nodules of various sizes. Residual connections were also included to reduce false detections, where datasets were used to show improved sensitivity using various lesion sizes. How- ever, this framework is limited to detecting nodules, not classifying them.

   Saha et al. [3] created Lung-AttNet, an attention-based convolutional neural network. In order to identify discriminative areas in CT scans, their approach uses lightweight attention modules. Additionally, predictions are interpreted us- ing visualisation methods like Grad-CAM and LIME. When compared to conven- tional CNN designs, the experimental results show better classication perfor- mance. However, real-time deployment is dicult due to the lack of localisation and the higher processing burden.

   Shakir et al. [4] introduced a radiomics-driven Bayesian inversion approach for malignancy and stage prediction. The system extracts from quantitative imaging biomarkers, including shape irregularity, entropy, and the surface fea- tures, which are often used for probabilistic inference. Results indicate reliable stage estimation under the controlled conditions. Nevertheless, this approach is highly depends on accurate segmentation and can degrade when boundaries are unclear.

   Javed R. [5] conducted a systematic mapping study of deep learning for lung cancer classication, structuring their analysis into the three phases: Study Preparation, Conduct of Study, and Analysis and Results. The study was pre- ceded by a comprehensive Systematic Literature Review (SLR) to classify method- ologies and identify deep learning application trends for a lung cancer detection. It incorporated with state-of-the-art practices, focusing on newer advancements in CNN Architectures.

3. System Architecture

   Lung CT images from the IQ-OTHNCCD dataset are prepared for the CNN input by resizing, normalizing, and formatting them according to the suggested system. Random oversampling balances the class distribution before splitting the data into validation, testing sets and training.A multi-layer convolutional neural network uses convolution, pooling, and dropout operations to extract spatial features. After attening, the learned features move to fully connected layers for the classication[11]. The models performance is actually evaluated during training with the Adam optimizer and categorical cross entropy loss, which allows seamless communication between backend inference modules and frontend interfaces for the visualization and reporting.

   Fig. 1. Workow of proposed methodology

4. Dataset

   The IQ-OTHNCCD Lung Cancer Dataset is used in this study to classify lung cancer using images obtained through a CT scan. This dataset includes grayscale images of CT scans classied into three types: benign, malignant, and normal,

   as classied by experts. This dataset has variations in intensity, contrast, and structure, which are representative of real-world scenarios. The IQ- OTHNCCD Lung Cancer Dataset is used in this study to classify lung cancer using images obtained through a CT scan. This dataset includes grayscale images of CT scans classied into three types: benign, malignant, and normal, as classied by ex- perts. This dataset has variations in intensity contrast, and structure, which are representative of real-world scenarios.

5. Data Preprocessing

   All images were modied to a standard resolution of 128 x 128 pixels and then reshaped into a single-channel format for use with a convolutional neural net- work. The pixel values were also normalized to a range of [0,1] by dividing by

   255. The class labels were converted into parts like one-hot encoded vectors for the multiclassication.

   Fig. 2. Dataset images

6. Data Sampling and Dataset Splitting

   To tackle the problem of imbalanced data in the benign, malignant, and nor- mal classes, the Random Oversampling (ROS) technique was employed on the attened data to attain a balanced data distribution. The data is then reshaped to the original image form and split into validation, training and test sets. The data is initially split into 70% for testing and the remaining 30% for training, with 20% of the 30% being used for validation

   Fig. 3. Original class distribution

   Fig. 4. Class distribution after Random oversampling (ROS)

7. Model Architecture

   Convolutional Neural Network: A convolutional neural network is used to classify lung CT scan images into benign, malignant, and normal classes. The CNN is trained to eectively learn spatial and structural features from grayscale CT scan images of size 128 × 128 × 1. The CNN architecture includes pooling, dropout, convolutional and fully connected layers.

   Fig. 5. Proposed CNN Architecture

   Input Layer: The input layer takes pre- processed grayscale CT scan images of size 128 × 128 × 1. This is a standard size, ensuring consistency throughout all images.

   First Convolutional Block: The rst convolutional layer employs 32 lters of the size 3 × 3 with ReLU activation function. The goal of this layer is focused on learning low-level characteristics such as contours, edges and basic texture features present in lung CT scan images. A max pooling layer is used to reduce spatial dimensions of size 2 × 2 while preserving essential features. In order to avoid overtting, a dropout layer is included with a dropout rate of 0.01 by randomly dropping a few neurons while training. Second Convolutional Block: The second convolutional layer uses 64 lters of size 3 x 3 with ReLU activation function[1]. This layer is used to learn more abstract features by utilising the features learned in the previous block. A 2 x 2 max pooling layer is used to reduce the feature map size. A dropout layer is then used with a dropout rate of 0.01. Third Convolutional Block: The third convolutional block uses the same parameters as the previous block: 64 lters with ReLU function. This block is used to learn complex features that will help the network distinguish between dierent lung tissue types. The max pooling and dropout layers are again used to avoid overtting.

   Flattening Layer: The feature maps learned in the convolutional blocks need to be crushed into a single dimensional vector. This is done in the attening layer. Fully Connected Layer: A dense layer which contains 128 neurons and ReLU anchor activation function is useful to learn the features from the attened data. This is the nal

   feature extraction block. A dropout layer of 0.01 dropout rate is used to avoid overtting.

   Output Layer: This layer uses three neurons, one for each of the three classes. Softmax activation function is used in this layer so the network can assign a condence value to each prediction.

   Training Conguration and Batch Processing: The CNN model is trained with the help of Adam optimiser, which adaptively adjusts the step size to ensure stable and ecient convergence. Tier cross entropy is used as the loss function, as the task include multi-class classication. Training is arranged for 10 epochs with a batch size of 32, meaning that 32 images are processed simultaneously in each iteration. Batch-wise training improves computational eciency, stabilises gradient updates, and allows eective utilisation of available memory. Model performance is monitored using validation data during training.

8. Principal Component Analysis

   In this context, PCA is used as a dimensionality reduction technique to extract important features using CNN feature representations. This reduces the dimen- sions while preserving the most important features in the feature space. This makes the feature space computationally ecient and reduces the redundancy. PCA[9] is a technique for dimensionality reduction that encompasses a series of steps such as feature extraction, covariance matrix calculation, eigenvalue decomposition, transformation and selection of principal components.

   The rst move in PCA is feature extraction using the CNN model. Here, the CNN feature

   extractor is used to generate high-dimensional feature representations for the input CT scan images. The CNN feature extractor is composed of all the layers in the CNN architecture except for the last two fully connected layers[3]. This is because the spatial information is preserved in the last two layers during CNN training. Once the features are extracted for the training and test datasets, they are used as input to the PCA algorithm. PCA normalises the feature vectors. This makes achieved by subtracting the mean value of each feature for every data point in the feature space. This is to maintain the feature space at a neutral position in the coordinate system, i.e., at the origin. This is because any existing bias in the feature space due to pixel intensity may be reduced.

   After mean aligning, the covariance matrix is calculated. The covariance ma- trix maintains connection between the various features, showing the relationship as they vary with respect to one another. High covariance between the features implies redundancy, while low covariance implies that the features are indepen- dent of one another. PCA[9] identies the direction in which the data shows maximum variance using the covariance matrix.

   To nd the main components, the covariance matrix is decomposed using the eigenvalue decomposition method, which gives the eigenvalues and the cor- responding eigenvectors. The eigenvectors are the directions in the feature space in which the variance of the data is maximum, while the eigenvalues are the magnitudes of the corresponding eigenvectors. The eigenvectors with the max-

   imum eigenvalues contain the maximum amount of descriptive data about the variance in the data.

   Principal component selection is the selection of the eigenvectors with the maximum eigenvalues, i.e., PCA selects the rst two components (n-components

   = 2) from the CNN features, which maps the data onto a space with two dimen- sions.

   Fig. 6. Confusion matrix

   Table 1. Model Comparison Table

   Model Accuracy (%) Precision (%) Recall (%) F1 Score (%)

   Naive Bayes	79.17	80.51	79.17	78.90
   Decision Tree	81.37	81.36	81.37	81.34
   KNN	83.15	83.46	83.15	83.24
   SVM	90.01	90.17	90.01	90.06
   LDA	90.86	91.26	90.86	90.94
   CNN (Proposed)	95.09	95.19	95.09	95.09

9. Results

   When we consider the performance of the suggested CNN classier for the lung nodule images, we can see that the accuracy, precision, recall, and F1-score are the key performance indicators. These results are demonstrated in the Evaluation Metrics Bar Graph. The accuracy of the model is around 95%, which is a clear

   indicator of good performance for the classication of images. This also shows that the CNN is successfully extracting useful spatial features from the images. Similarly, a precision of about 95% is an indicator of fewer false positive instances and better positive predictions. The recall is about 95%, which is a clear indicator that most of the malignant cases are being identied. This is a great performance, especially when considering the early diagnosis of lung cancer. The F1-score is also estimated at about 95%, which is a clear indicator of a good balance of precision and recall. This is true for all classes, and when the results are above 90%, it is a clear indicator of the robustness of the model.

10. Conclusion

    This paper presented a CNN-based system for classifying lung nodules using the IQ-OTHNCCD Lung CT dataset, dividing them into Benign, Malignant, and Normal classes. After precise preprocessing, using a Random Over Sampling method, and a well-congured CNN, the system achieves remarkable perfor- mance: 95.09% accuracy, 95.19% precision, 95.09% recall, and 95.09% F1 score. Its high recall rate is evidence of its eectiveness in a medical context, where early identication of malignant nodules is critical, while its strong F1 score is evidence of fair classication performance. Thus, this system is a highly eective tool to assist radiologists in lung cancer diagnosis, showing considerable potential for future upgrades.

    Fig. 7. Evaluation Metrics

    References

    1. Liu, K.: Stbi-yolo: A real-time object detection method for lung nodule recognition. IEEE Access 10, 7538575394 (2022)

    2. Nguyen, C.C., et al.: Pulmonary nodule detection based on faster R-CNN with adaptive anchor box. IEEE Access 9, 154740154751 (2021)

    3. Saha, C., et al.: Lung-attnet: An attention mechanism-based CNN architecture for lung cancer detection with federated learning. IEEE Access (2025)

    4. Shakir, H., et al.: Radiomics-based Bayesian inversion method for prediction of cancer and pathological stage. IEEE Journal of Translational Engineering in Health and Medicine 9, 18 (2021)

    5. Javed, R., et al.: Deep learning for lung cancer detection: a review. Articial In- telligence Review 57(8), 197 (2024)

    6. Li, Y., et al.: Machine learning for lung cancer diagnosis, treatment, and prognosis. Genomics, Proteomics & Bioinformatics 20(5), 850866 (2022)

    7. Saleh, A.Y., et al.: Lung cancer medical images classication using hybrid CNN- SVM (2021)

    8. Swain, S., et al.: Innovative approach for early detection and risk assessment of lung cancer using machine learning techniques. In: International Conference on Intelligent Computing and Sustainable Innovations in Technology (IC-SIT), IEEE (2024)

    9. Bhargav, S.P., et al.: Impact of PCA on lung cancer dataset classication: A com- parative analysis of machine learning models. In: International Conference on Ad- vances in Data Engineering and Intelligent Computing Systems (ADICS), IEEE (2024)

    10. Nair, S.S., Devi, V.N.M., Bhasi, S.: Enhanced lung cancer detection: Integrating improved random walker segmentation with articial neural network and random forest classier. Heliyon 10(7) (2024)

    11. Nimmagadda, S.M., et al.: Lung cancer prediction and classication using machine learning algorithms. In: International Conference on Expert Clouds and Applica- tions (ICOECA), IEEE (2024)

    12. Nageswaran, S., et al.: Lung cancer classication and prediction using machine learning and image processing. BioMed Research International (2022)

    13. Das, S., Majumder, S.: Lung cancer detection using deep learning network: A com- parative analysis. In: Fifth International Conference on Research in Computational Intelligence and Communication Networks (ICRCICN), IEEE (2020)

    14. Slack, D., et al.: Explaining machine learning models with interactive natural lan- guage conversations using TalkToModel. Nature Machine Intelligence 5(8), 873 883 (2023)

    15. Shravya, V., et al.: Encouraging discriminative attention through contrastive ex- plainability learning for lung cancer diagnosis. IEEE Access (2025)