This study exploits a color image dataset, LC25000 [37], released by James A. Haley Veterans’ Hospital, Tampa, Florida, United States of America (USA). LC25000 contains histopathological color images of cancerous colon and lung cancers. However, this research only utilizes cancerous and benign lung tissue images. Lung carcinomas are the most predominant source of invasive cancer and a major cause of high mortality rates in America [38]. In the image acquisition phase of the LC25000 database, the rules of health insurance portability and accountability act (HIPAA) were completely followed, and the 750 color lung tissue images, including lung squamous cell carcinomas (lung-scc), benign lung tissues (lung-n), and lung adenocarcinoma (lung-aca), were acquired from microscopy slides [39].

Following this, an augmentor package in python programming language is used for the image augmentation of the pathology images [40]. The dataset with respect to lung cancer tissue images is expanded to 15000 color images, comprising 5000 lung tissue images for each class, including lung squamous cell carcinomas, benign lung tissues, and lung adenocarcinoma. Later, the images were resized to 224 × 224 pixels, and the labels were encoded using a one-hot encoder style for our multi-classification task. The data used in this study from the LC25000 is publicly available, fully de-identified, validated, and HIPAA compliant. LC25000 is widely used and acknowledged for its effective augmentation-driven expansion but also one of its core limitations due to limited patient diversity. Therefore, we would like to clarify that the proposed research and its promising outcomes reflect the enriching power of the proposed framework in pattern recognition within this benchmark, not immediate clinical generalization.

Preprocessing plays an important role in refining the data crucial for correct classification. For image enhancement, two effective filtration techniques were used in this study, including CLAHE and MF. The most common challenge in the benchmark histogram equalization strategies is the amplification of noise during the enhancement process [41]. However, CLAHE showed promising outcomes in enhancing the quality of the images by improving contrast without significant noise amplification. The prime aim of CLAHE is to prevent darkness and cover bright regions in the image. It works by separating the contextual regions of the image and employing histogram equalization to each region. This region-specific strategy enables precise contrast enhancement for each region based on the specified needs. To address the overamplification concerns, CLAHE utilizes a clip threshold on the histogram equalization procedure within the regions. This threshold Tclip, controls contrast adjustment by lowering the intensification effects differently for each region. The process includes the conversion of images into lightness, green-red, blue-yellow (LAB) color space, to detach the L∗channel, which carries the most useful information for contrast adaptation. The process of applying CLAHE to this channel is given in Eq. (1). (1)L∧=CLAHE(L∗)=L∗(TclipTmax−1)×(L∗−L¯)where L∗ refers to the real brightness value, L∧ corresponds to the adjusted brightness value, L¯ is the average brightness of the bordering regions, Tclip is the threshold clip limit and Tmax is the uppermost histogram score allowed per region.

Additionally, a median filter is exploited in this study for removing the noise in the histopathological images. Median filters have shown promising results in the noise removal process of medical imaging datasets. MF lowers image noise by substituting outlier pixel values with the median of its adjacent values. The MF working process depends on a single parameter, which is the size of the neighborhood window.

Fig. 3 shows the quality difference between the original and the enhanced histopathology image using CLAHE and MF. We have performed critical hyperparameter tuning associated with CLAHE and MF to obtain the best possible image enhancement. Fig. 4 shows the sample of enhanced images under different parameter values. The optimal hyperparameter values were a window size of 5, a tile size of 8, and a threshold of 40.

Existing medical image processing approaches often struggle to strike a balance between local and global feature extraction. Therefore, the suggested CAFNet is critically designed to overcome these constraints. The detailed pseudo code and the graphical overview of the CAFNet are given in Table 2 and Fig. 5, respectively. Additionally, Table 3 presents an extensive architectural insight into the proposed network with respect to the positioning and repetition of the blocks in the designed network. After the effective preprocessing of the data, the images will undergo multiple MIBConv layers for the initial feature extraction process. The MIBConv layer utilizes depthwise separable convolutions, which are both effective and computationally inexpensive operations. Afterwards, the network is divided into two different phases. The first phase includes the division of feature maps into four patches. Thereafter, each patch separately moves to SARL and CARL techniques for spatially localized feature learning (SLFL). These modules are carefully designed to focus on different regions in the image and highlight the most prominent ones. The outcomes of these attention modules are then passed to a global average pooling (GAP) operation for compressing and summarizing the information into a manageable style. Subsequently, the second phase of the network emphasizes holistic feature extraction (HFE) by focusing on the entire image instead of patches. The second phase uses the same attention modules but on a broader scale. These two pathways helped the network to capture both the local and global features through two independent learning schemes. Following this, both sets of features are concatenated and passed to the specified fully connected layers for further refining of the features and representation process. After the complete learning scheme of the network, it uses SoftMax activation operation for classification purposes. This extensive configuration of the modules sets CAFNet as a potential solution in medical image classification, providing noteworthy gains in accuracy and efficiency.

The transparency [42] of the solution is critical in healthcare disciplines for building trust in AI-based diagnostic tools. For that reason, we have exploited the concept of explainable AI to produce a transparent and human-understandable solution for interpretable lung cancer analysis. Explainable AI (XAI) [43] provides interpretability by offering key findings on the decision-making process and qualitative insights into the solution. The transparency of the proposed network is analyzed through two different XAI-based approaches, mainly Grad-CAM and LIME, which show promising results in elaborating the decisive behavior of the AI-based models. Both interpretability schemes helped to identify the most prominent regions within the images and contributed mostly to the specified prediction of the model. Table 4 shows the complete working procedure of Grad-CAM in explainable visualization of the network predictions [44]. Similarly, the schematic diagram of LIME-based explainable AI is given in Fig. 9. The optimization objective of LIME involves two loss functions. The key objectives in the optimization procedure of LIME are (I) to reduce the difference between simple and complex models, (II) to control the complications of the simple model for providing interpretability. The prime aim of LIME [45] is to ensure interpretability, transparency, and local faithfulness by reducing the first loss term and maintaining the second loss as minor enough to be understandable for a human.

PADAM is a modified version of Adam, which bridges the gap between convergence and generalization capability [46] of the optimizer. It presents a new factor Pt, to modify the degree of adaptivity in the optimization strategy. The pseudo-code-based mathematical intuition behind the PADAM is presented in Table 5.

In Table 5, bt represents the resulting stochastic gradient and e^t denotes the moving average of 2nd degree momentum for the computed stochastic gradient. The core differentiator depicted from Table 5 with respect to standard Adam [47] is the addition of “partially adaptive term (Pt)” by the 2nd degree momentum:(15)ωt+1=ωt−βt⁢nte^tPt,H⁡e⁢r⁢e⁢e^t=max⁡(e^t−1,e^t)

PADAM transforms the standard Adam optimizer by varying the 2nd momentum explored in the adaptive learning rate. The overall update rules for both standard Adam and suggested PADAM are given below:(16)ωt+1=ωt−σ.nte^t+ε (Adam)where, nt and et are 1st and 2nd order momentum estimates, σ is the learning rate (LR) and ε is a constant term. (17)ωt+1=ωt−σ.nte^tPt+ε (PADAM)

In Eq. (17), e^t is raised to the power of Pt (0≤Pt≤1), which makes the learning rate partially adaptive. Eq. (16) uses e^t for an adjustable learning rate, which occasionally resulted in a poor generalization outcome. However, PADAM introduces a partially adjustable term (Pt) which tunes the supremacy of et. This addition can decrease the adaptiveness of learning trends to enhance the generalization properties. Through the control of adaptiveness, PADAM intends to bridge the gap between convergence and generalization. Exploration of partially adaptive term (Pt) permits for a more supervised convergence procedure, evading the excessive adaptiveness which leads to obstructed convergence.

The proposed CAFNet is executed with a standard Adam optimizer and four different adaptivity terms of the PADAM optimization approach under the same setting to attain the optimal degree of adaptivity with respect to the learning rate. Table 6 shows the brief performance analysis of the CAFNet with Adam and PADAM different Pt values. Initially, the CAFNet was trained with a standard Adam optimizer for 20 iterations, which resulted in an accuracy of around 89%. Later, the network was executed with a PADAM (Pt=0.25) for 20 epochs, which showed satisfactory classification performance by achieving a test loss and accuracy of 0.1523 and 0.9390. Following this, the degree of adaptivity further increased from 0.25 to 0.5, which resulted in an improved diagnostic performance. With Pt=0.5, the CAFNet attained a substantial test loss and accuracy of 0.7779 and 0.9717. Similarly, the performance in terms of other evaluation metrics like precision, recall, and f1-score for each category of cancer was noteworthy, showing the generalizability of the proposed approach. Afterwards, the degree of adaptivity is further increased to 0.75, which showed the decremental trend in the learning behavior as compared to previous adaptivity factors. With Pt=0.75, the proposed network saw a significant drop in predictive performance by attaining an accuracy of 0.9417. The path towards the divergence behavior was quite visible through this degree of adaptiveness in the optimization process. However, further analysis verifies the performance of CAFNet with Pt=1.0, which provided us with a clear idea about the complete divergence in the optimization process, as shown in Table 6. The variations in the learning behavior of the proposed network with different Pt values are presented through the learning curves given in Fig. 10. A similar trend of adaptiveness in the PADAM optimization can also be observed through the AUC-ROC presented in Fig. 11, which showed the superiority of the proposed CAFNet with PADAM (Pt=0.5) as compared to its counterparts. To sum up the discussion on the identification of an optimal degree of adaptiveness in the optimization process, Fig. 12 presents the confusion matrix at four different Pt values. Hence, we concluded that the proposed CAFNet model showed optimal classification performance with PADAM (Pt=0.5).

After the performance analysis through the on-hold validation approach, we have further performed a 5-fold validation technique by splitting the dataset into multiple training and testing sets to prove the effectiveness, reliability, and generalizability of the proposed model. This validation is quite useful in evaluating the predictive performance of the deep learning networks, particularly in healthcare applications. Table 7 shows the detailed performance overview at different folds and iterations, which provides crucial insights into the effectiveness of CAFNet as a classifier for NSCLC classification. The model demonstrates superior classification proficiencies by showing a consistent average accuracy of around 0.95. For each fold, the performance of the CAFNet was observed at four different iterations with respect to both loss and accuracy. Additionally, the mean and standard deviation (std) of folds were also reported for each iteration stage. The model showed consistent accuracy across iterations, proving the enriched convergence of the proposed approach. Similarly, the low std across loss and accuracy displays the steadiness and robustness of the proposed model for the task at hand. Table 8 shows the progressive rise in the classification accuracy with the inclusion of each component in the proposed framework. It is worth noting that the addition of each component contributes significantly to the classification performance of the proposed CAFNet. However, the comprehensive analysis demonstrated that the integration of suggested components is critical for obtaining optimal results.

This study exploits two different approaches to explainable AI, such as LIME and Grad-CAM, to provide crucial insights into the decisive behavior of the proposed network. Despite the emergence of AI, the transparency of the deep networks is still the key point of concern, particularly in sensitive applications like healthcare. It’s necessary to provide the human understandable explanations on the model’s qualitative insights and decision-making process. Fig. 13 presents a LIME-based explanation of the CAFNet’s prediction for the given histopathological image. The image shows the original and the highlighted map, providing a super pixel-level explanation of the given prediction. The regions highlighted in yellow circles indicate the most prominent and influential regions behind the specified prediction of lung squamous cell carcinoma. Additionally, the feature contribution plot in the explanation offers more crucial insights on the approximated positive weights, showing the contribution of these superpixel-based regions. Through this explanation, we can conclude the most significant aspects of the images and the decision behavior of the proposed model. Similarly, Fig. 14 shows the Grad-CAM-based explanation of the CAFNet’s decision-making process. The Grad-CAM heatmap dominantly highlights the most contributed spatial regions of the given histopathology image with respect to the lung squamous cell carcinoma. The red/yellow color scheme on the heatmap shows the most contributed regions, and other cooler colors indicate the regions with minimal or no contributions. The Grad-CAM overlay provides an umbrella view of the decisive model by mapping the heatmap on the original image and highlighting the most significant regions for human understandability.

To verify the discriminative abilities of the proposed network, this study utilized t-SNE driven two-dimensional visualization approach [48]. Initially, the principal component analysis (PCA) is used to reduce the dimensionality of the learned feature vectors from the final dense layers, which is better for the stabilization and the clear visualization of the high-dimensional space. Afterwards, the 2D embeddings are created for the projection of the learned features. Fig. 15 shows the detailed t-SNE visualization of the proposed CAFNet with different Pt values. It can be observed that the proposed network along PADAM (Pt=0.5) shows the best performance by producing very compact and separated clusters as compared to its counterparts. These qualitative analyses clearly demonstrate the enriching capabilities of the CAFNet in increasing the inter-class and decreasing intra-class relationships.

To overcome the accessibility concerns, we have deployed our proposed CAFNet on a web-based user interface using a Python library Gradio [49]. The Gradio-driven web interfaces are helpful in bridging the gap between academic research and clinical usage. Through the designed web-based interface, the patients, researchers, and healthcare professionals can easily access our proposed solution for receiving real-time predictions on their histopathological domain images. The designed interface is user-friendly, equipped with real-time interaction leading to more reliable and fast response. With the help of this smart healthcare system, we would be critically monitoring the performance of the proposed network on the diverse histopathology data for the NSCLC classification. Fig. 16 shows our designed web-based interface using the proposed CAFNet architecture, requiring a histopathology domain image from the user to provide the prediction, confidence in the given prediction, and the Grad-CAM-driven explanation on the given user input. This smart healthcare-based application will be extremely helpful for us to improve the performance of the model as per the needs and feedback of the users. Fig. 17 presents the overall information flow of the application from the given input to the final prediction by our proposed model.

The outcomes of the study were quite promising with respect to the successful development of a smart healthcare system for NSCLC classification through histopathology. Following this, we have compared the performance of the proposed network with the existing state-of-the-art models for NSCLC classification, given in Table 9. It is depicted that the CAFNet has shown superior performance by outperforming its counterparts in terms of the standard evaluation measures like accuracy (ACC), precision (PRS), Recall (RC), and f1-score (F1S). The dashes (- - -) present in the comparison table indicate the evaluation measures not reported in the existing works. Similarly, all the comparisons with the existing methodologies are reported at the benchmark level and are not protocol-matched. The proposed network has achieved an impressive accuracy of 97.17% for the task at hand. Additionally, the computational complexity of the network is also compared with the existing models in terms of the model’s architectural parameters (AP) in millions and model size (MS) in MB. From Table 9, it is quite evident that the proposed model is impressively efficient with respect to the computational complexity as compared to its counterparts. Hence, it was concluded that the proposed CAFNet serves as an accurate, efficient, and computationally inexpensive solution for the NSCLC classification task.