CNN-EWC: A continuous deep learning approach for lung cancer classification

1 Introduction

The field of medical imaging has experienced a transformative revolution with the introduction of deep learning techniques, particularly in the detection and diagnosis of lung cancer. Convolutional Neural Network (CNN) and advanced deep learning architecture have exhibited unparalleled capabilities in analysis of medical images, often surpassing or matching the performance of human experts [1,2,3].

In recent studies, the potential of deep learning to detect early-stage lung cancer has been demonstrated. Machine learning models have exhibited exceptional sensitivity in the identification of subtle radiological patterns that may evade traditional diagnostic methods. For pulmonary image analysis, researchers developed sophisticated neural network architectures. In a multitude of clinical datasets, these architectures have obtained detection accuracy that surpassed 95% [4,5].

In the detection of lung cancer, the primary benefits of deep learning are as follows: Enhanced feature extraction capabilities, the capacity to process complex multi-dimensional imaging data, the potential for early and more accurate diagnostic predictions, and reduced inter-observer variability in image interpretation [6,7].

Despite these developments, there are still substantial obstacles to overcome, such as the necessity for extensive annotated datasets, the generalizability of the model, and the capacity to adjust to the changing landscape of medical imaging technologies [8].

The fundamental challenge of knowledge retention and adaptation in neural networks is addressed by continuous or perpetual learning, which represents a critical frontier in machine learning.

Elastic Weight Consolidation (EWC) is a critical approach for preventing catastrophic forgetting, a phenomenon in which neural networks swiftly overwrite previously learned information when they are trained on new tasks [9,10,11,12,13].

EWC introduces a sophisticated regularization technique that effectively enables models to learn new information while maintaining previously acquired knowledge, preserving key network parameters during subsequent training iterations of the model. EWC facilitates more adaptive and stable learning processes by estimating parameter importance using a probabilistic framework [14,15,16,17,18].

Current deep learning models for lung cancer diagnosis encounter significant obstacles, such as catastrophic forgetting, reliance on extensive annotated datasets, inadequate generalizability, and instability in prolonged learning. Conventional models face challenges in preserving prior knowledge while accommodating new information, hence constraining their efficacy in continuous learning contexts.

The proposed CNN-EWC model mitigates these challenges by maintaining critical parameters via EWC, thereby facilitating knowledge retention and adaptation to new tasks. This approach reduces the need for extensive retraining, enhances generalizability across datasets, and maintains model stability over time. CNN-EWC offers a more efficient, flexible, and dependable deep learning framework for lung cancer diagnosis in practical medical applications. The main contributions of this work are summarized as follows:

• Involved in building effective techniques for lung cancer classification and detection with CNN enhanced with EWC.

• Model designed to tackle the problem of catastrophic forgetting in lifelong deep learning in which serious weights are retained at the time of incremental learning.

• The model was evaluated on histopathological images of cancerous tissues with high accuracy and functionality over varying datasets, proving its readiness for real medical diagnosis.

The remaining sections of the study are organized as follows, In Section 2, the related works are presented, while Section 3 introduces the proposed method. Section 4 presents the results and discussions; Section 5 shows the limitations of the proposed model and Section 6 concludes with a few suggestions for future research.

2 Related works

In the quickly advancing field of medical image analysis, deep learning methodologies, especially CNNs, have shown significant potential in the detection of lung cancer and the enhancement of diagnostic imaging. Several foundational investigations have utilized comprehensive image datasets, with significant research employing large-scale collections of histopathology images to train and validate advanced CNN architectures [19].

For example, the author presented study concentrates on the detection of lung cancer through the application of deep learning techniques, with a particular emphasis on CNNs. CNNs are utilized for the automatic classification and detection of lung cancer in histopathological images, thereby minimizing dependence on manual analysis. The research employs an extensive dataset of histopathological images pertaining to lung cancer. The CNN architecture is designed to extract features and recognize patterns associated with lung cancer. The model for the mentioned study attained an accurate score of 95%, illustrating its capability for precise classification. The findings indicate that CNNs can markedly enhance the early identification of lung cancer. Nonetheless, the limitation of the study highlights the critical role of deep learning in enhancing the efficiency and accuracy of diagnoses performed by pathologists. The manuscript lacks adequate information regarding data preprocessing, augmentation techniques, and computational demands, all of which are critical factors influencing model performance and scalability. The identified factors underscore the necessity for additional validation and refinement of the approach in subsequent research endeavors [20].

Another study investigates the application of ResNet50 architecture, CNN recognized for its efficacy in image classification tasks. The ResNet50 model has been developed to categorize lung cancer into various classifications utilizing histopathological images. The model demonstrated a notable accuracy of 98%, suggesting its capability for precise diagnosis.

The results exhibit certain limitations; external validation across diverse datasets is essential to confirm their applicability in real-world scenarios, the model’s performance needs to be assessed in clinical settings to determine its robustness and adaptability to variations in imaging equipment and patient populations [21].

Nayak et al. [22] introduced a method that combines an image preprocessing pipeline with spatial attention mechanisms to enhance the performance of the model. The model employs a neural network architecture (NetB3), which improves its capacity to concentrate on pertinent regions of the image. The model attained a classification accuracy of 99.4%, illustrating its efficacy in identifying lung cancer. Nonetheless, the study’s limitations fail to account for potential class imbalances within the dataset, which may impact the accuracy of the model. The absence of information regarding the computational resources necessary for training is noteworthy, as this could pose a limitation for practical applications. Finally, the potential for preventing overfitting has not been addressed, particularly in the absence of external validation or cross-validation on data that has not been previously encountered.

Continuous deep learning, often referred to as continual or lifelong learning, represents a developing area within artificial intelligence. In this domain, models are designed to acquire new tasks incrementally while preserving knowledge from earlier tasks. In contrast to conventional machine learning approaches that typically operate under the assumption of a static dataset, continuous learning effectively tackles situations where data are incrementally accessible over time. Methods such as EWC are effective in addressing catastrophic forgetting, a phenomenon where models tend to lose previously acquired knowledge when exposed to new tasks. The limitations of current approaches include challenges in ensuring models can generalize across diverse tasks and datasets, as well as the difficulty in adapting to rapidly evolving data streams without significant retraining [23].

In 2023, Li and colleagues presented a Continuous learning system (CLS) that utilizes deep learning in conjunction with optimization and ensemble methods to improve accuracy in medical diagnoses. The study employed ultrasound imaging of breast masses, involving 629 masses and 2,235 images from 561 cases, to train the model through 6 stages aimed at differentiating between benign and malignant tumors, as well as recognizing specific pathological types and diseases. The CLS underwent evaluation through the analysis of 7 independent datasets and was benchmarked against the assessments of 21 physicians. At the sixth training stage, the diagnostic performance of the system exceeded that of 20 medical professionals. This innovative approach shows potential for accurate breast mass diagnosis and could be adapted for intelligent diagnostics in other organs, thereby advancing the application of artificial intelligence in precision medicine. However, there are limitations to the study including physicians utilized only one dataset for testing, which comprised only 629 breast masses. Additionally, the diagnostic effectiveness of uncropped images was superior to that of cropped images, but this finding requires multi-center verification [24].

The study in 2023 conducted by Wang and colleagues proposed a novel deep learning model named CoroTrans-CL that was developed for the purpose of recognizing coronavirus infections in CT and chest X-ray images. The model utilizes the Swin Transformer architecture and integrates both regularization-based and rehearsal-based continuous learning strategies. The evaluation of CoroTrans-CL was conducted using the CL-COVID dataset, comprising 954 CT and 470 X-ray images for training, through joint training and continual learning experiments. In the joint training setting, the model attained an accuracy of 95.34% and a precision of 97.18%, demonstrating a performance enhancement of 16–20% compared to alternative models. In the context of continual learning, CoroTrans-CL achieved an accuracy of 83.40%, surpassing other models by roughly 30%. While the results underscore its potential in coronavirus image recognition, its adaptability to other datasets and medical imaging tasks requires further investigation; however, the proposed methodology underwent evaluation utilizing a designated dataset comprising CT and X-ray images pertaining to coronavirus infections. Its efficacy in relation to alternative datasets or different medical imaging applications may exhibit variability. The study does not address the model’s ability to generalize across different types of imaging devices or real-world clinical scenarios, where variability in image quality and data distribution might affect its performance [25].

In 2020, a study conducted by Yujiang implemented continuous learning methodologies to enhance forecasting accuracy in regional energy markets. This was achieved utilizing a dataset that included 1 year of 15 min resolution data sourced from 10 renewable energy generators and 10 local consumer plants. The evaluation encompassed methods including Learning without Forgetting, EWC, online-EWC, and synaptic intelligence. The experimental findings indicated that EWC and online-EWC significantly mitigated catastrophic forgetting, ensuring a minimal forgetting ratio. However, these methods increase training time depending on the scenario. Limitation of study is that the experiments are run on data from a small power grid, and the findings of this article can be applied to any level of the power grid where catastrophic forgetting can occur and EWC and online-EWC algorithms, while effective in reducing the forgetting ratio, may increase the training time compared to other algorithms [26].

3 Proposed model

This methodology facilitates the model’s ability to acquire new tasks while mitigating the risk of catastrophic forgetting, which is a significant concern in the realm of continuous learning. The Fisher Matrix is employed by the model to determine which parameters are essential for previous tasks, imposing penalties on substantial alterations to those parameters in subsequent training sessions. The proposed model is illustrated in Figure 1.

Figure 1

The proposed EWC model. Source: Created by the authors.

3.1 Loading lung cancer dataset

The LC25000 datasets [27] are used, which contain five classes of histopathology images (colon and lung cancer), but only three classes for lung cancer are used. There are 5,000 images of lung benign tissue (n), 5,000 images of lung adenocarcinoma (ACA), and 5,000 images lung squamous cell carcinoma (SCC), every image in the dataset is in jpeg format and has a resolution of 768 × 768 pixels. Table 1 illustrates the three types of lung cancer that are used in the proposed model.

Table 1

Cancer type and class name of the used dataset [27]

Cancer type	SCC	n	ACA
Sample image for each type
Class name	Lung_SCC	Lung_n	Lung_ACA
Number of images	5,000 images	5,000 images	5,000 images

4 Results and discussion

In this section, the results of the proposed model are discussed, in which the dataset was used [27]. 13,500 images were trained at three different times and a different number of images. The obtained results are shown in Table 4.

The results in Table 4 showed that as the quantity of images per class diminishes in later tasks, the performance remains stable. This demonstrates the model’s resilience to a decrease in the training samples, particularly for Lung_n, which reliably attains perfect or near-perfect outcomes. Also, the model demonstrates exceptional classification performance across tasks, with accuracy increasing over time (98% to 99.4% to 99.6%).

The efficacy of continual learning strategies such as EWC is underscored by the gradual decrease in misclassifications. Finally, the results underscore the effectiveness of EWC in preventing catastrophic forgetting. The model ensures knowledge retention by maintaining high performance on previous tasks, despite the addition of new ones. Figure 2 illustrates the confusion matrix for the proposed model for three tasks.

Figure 2

Confusion matrix for the proposed model. Source: Created by the authors.

A lung disease classification system’s performance on three separate tests is shown by these confusion matrices. For three lung conditions (lung_ACA, lung_n, and lung_SCC), the matrices compare predicted labels to actual diagnoses. With little interactor misclassifications and high diagonal scores (578 for ACA, 621 for normal tissue, and 555 for SCC), the model performed admirably in Task 1. With 430,471, and 431 correct predictions in Task 2, performance is little lower but still respectable. Either the dataset is more difficult, or the model is less successful for this work, since work 3 shows much lower performance across all classes (289, 280, and 324 correct predictions). Overall, the model does very well on Task 1, very well on Task 2, and with less performance on Task 3, but it still manages to classify all three kinds of lung tissue reasonably well.

The results of the accuracy and loss rate are shown in Figure 3.

Figure 3

The proposed model’s loss and accuracy. Source: Created by the authors.

In the first task, the training was stable at the end, where the hyper parameter (epoch = 20) was determined and the training stopped and was stable at (epoch = 18). Also, the percentage of losses was very small. The scatter in the first task starts from 70 and 98 that it took some time to train because it was the first time, and the number of images was large. Loss in Tasks 2 and 3 is considered good because in the end, it settled and stopped at epoch 18. Finally, the three tasks indicate that training accuracy and validation accuracy are close together, which indicates that the images were trained well.

Additional extra tests were conducted on the model to ensure that the images were classified correctly. Two extra tests were conducted, the first was using 1,500 images that the model had not been trained on, the results obtained are shown in Table 5. For the second one, images were taken from the first task and tested on the final obtained model to ensure that the model worked correctly and recognized the images that were trained on before, the obtained results are shown in Table 6.

The model exhibits significant robustness and accuracy across diverse dataset sizes, reflecting its strong performance in this classification task. Even for lesser datasets (100 or fewer images), accuracy metrics reach a plateau, indicating that large datasets may not be required for this task. Also, the model exhibits effective scalability with dataset size, attaining expedited processing times for smaller datasets without compromising accuracy.

The processing duration diminishes with a decrease in the number of images, varying from 0.16 s for 1,500 images to nearly 0 for 10 and 3 images. Finally, Lung ACA exhibits minor fluctuations, demonstrating near-perfect yet not absolute accuracy for extensive datasets.

The model exhibits impeccable performance, achieved completeness in all three categories. This suggests that the model demonstrates a significant level of efficacy for this dataset. The equitable distribution of images across the various classes may have played a significant role in achieving this elevated accuracy, given that no individual class predominates within the dataset. The classification process occurs at an impressive speed of 0.01 s, making it highly suitable for real-time applications.

Finally, a comparison with previous works is shown in Table 7, at which the CNN-EWC model attains impressive accuracy of 99.6%, illustrating its superiority compared to alternative methods for this classification task and the models demonstrate a steady enhancement in accuracy as they evolve from conventional deep neural networks (DNNs) to advanced architectures such as ResNet50, NetB3, and CNN-EWC.

5 Limitations of the proposed model

Current deep learning models for lung cancer diagnosis encounter significant obstacles, such as catastrophic forgetting, reliance on extensive annotated datasets, inadequate generalizability, and instability in prolonged learning. Conventional models face challenges in preserving prior knowledge while accommodating new information, hence constraining their efficacy in continuous learning contexts.

6 Conclusion and future work

The proposed CNN-EWC model has demonstrated continuous deep learning effectively with the ability to correctly classify images based on the dataset that the model was trained on at different time points. This work is based on providing a suitable image classification model that can classify correct images with high classification efficiency.

Contemporary deep learning models for lung cancer diagnosis encounter obstacles like catastrophic forgetting, dependence on extensive annotated datasets, restricted generalizability, and instability over extended learning periods. These constraints impede their efficacy in continuous learning applications. The proposed CNN-EWC model addresses these concerns by maintaining essential parameters and improving flexibility; nonetheless, additional validation is required to confirm its resilience across various medical datasets.

The research contributions for EWC algorithm were used to preserve important information that can be trained on later. In creating the deep learning model, the CNN algorithm is used to extract features, while the EWC algorithm is used to store parameters. It also eliminates the time problem.

The presented contribution is a CNN model with very high classification accuracy that enables the extraction of the most important image features to the level of 99.6%. This work requires entering a set of image data from time to time and training the model on it and extracting important information from the images and storing it and then retrieving it when other data arrive.

In this study, the practical advantage of the model was tested on a set of images that the model was trained on for the first time and the results were very high, reaching 100%, and it was tested on images that the model was not trained on and achieved high performance as it reached 100%.

Finally, the principle of continuous learning using the EWC method has proven its effectiveness in continuous training in terms of accuracy, speed, saving time, and correct classification of images.

Future work investigating the expansion of the suggested CNN-EWC architecture should accommodate more intricate multi-task scenarios including additional cancer kinds and larger datasets. Furthermore, including additional continual learning methodologies may further improve adaptability and efficacy in practical clinical application.