Integrating central-slice anatomy and MRF-mapped radiomics dynamics for lung adenocarcinoma subtyping

Introduction

Lung diseases, particularly lung cancer, represent a significant global health challenge. Lung cancer is the leading cause of cancer-related mortality, accounting for approximately 20% of all cancer deaths (1,2). Lung cancers are among the most prevalent and deadly pulmonary conditions, placing immense strain on healthcare systems worldwide (3). The rising incidence of these diseases is driven by risk factors such as smoking (4), air pollution (5-7), and genetic predisposition (8,9), with the overall burden expected to increase further. Early detection and intervention are crucial for improving patient outcomes—especially in lung cancer, where early-stage lesions often manifest as small pulmonary nodules. Computed tomography (CT) (10-12) has emerged as the most effective modality for screening lung cancer and pulmonary nodules (9,13,14). Its high spatial resolution facilitates clear visualization of subtle signs—such as lobulation, calcification, and vascular clustering—making CT superior to traditional chest radiography. Although most pulmonary nodules are benign, their potential for malignant transformation underscores the importance of early detection.

In recent years, deep convolutional neural networks have demonstrated strong effectiveness in medical image diagnosis (15-17). Wang et al. (18) proposed TransUnet for accurate automated lung nodule classification, focusing primarily on distinguishing malignant from benign lesions and providing insights applicable to broader lung cancer tasks. Yu et al. (19) introduced ETMO-NAS, an efficient two-step multimodal one-shot neural architecture search method for positron emission tomography (PET)/CT-based nodule classification, achieving superior performance with substantially fewer parameters than conventional CNN or NAS models. Lin et al. (20) further developed a hybrid model integrating deep learning, radiomics, and clinical variables to classify lung nodules into benign or malignant categories, refine pathological subtypes, and assign Lung-RADS grades using a stacked ensemble of modified 3D CNNs, radiomics signatures, and clinical features.

Building on these advancements, 3D-CNNs have shown particular strengths in capturing tumor morphology, spatial context, and volumetric structure, leading to improved diagnostic performance in various clinical scenarios (21,22). For example, Han et al. (23) trained a 3D-CNN to differentiate active pulmonary tuberculosis (APTB) from community-acquired pneumonia (CAP), achieving high accuracy in distinguishing these clinically overlapping conditions. Mikhael et al. (24) developed Sybil, a low-dose CT-based model capable of predicting long-term lung cancer risk, which achieved a high 1-year area under the curve (AUC) that declined over extended follow-up. Despite these advances, current diagnostic frameworks—whether radiomics-based, deep learning-based, or hybrid—often emphasize isolated morphological or textural features while overlooking the dynamic changes occurring across adjacent CT layers. This omission is crucial, as radiologists rely heavily on slice-by-slice evolution of lesion density, margins, and internal architecture when assessing progression from pre-invasive to invasive adenocarcinoma. Furthermore, many existing models lack mechanisms that explicitly integrate such dynamic progression into the computational pipeline, thereby limiting their alignment with clinical diagnostic reasoning.

This study addresses these gaps by aligning our analysis with the clinical diagnostic workflow used by radiologists. By concentrating on key slices that highlight lesion features and adjacent tissues, we capture both the static appearance of lesions and their dynamic evolution across multiple layers. Leveraging deep learning techniques such as BiLSTM and CNN + MLP, we aim to improve diagnostic accuracy by modeling lesion progression, thereby enhancing detection and classification of lung diseases. Our approach not only mirrors the clinical emphasis on the most diagnostically relevant slices but also tracks lesion changes across different layers, providing a more comprehensive and clinically aligned solution for lung disease diagnosis.

Methods

This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the ethics committee of Shanghai University of Medicine and Health Sciences (No. 2019-ZYXM-04-321102****), and the written informed consent from patients was waived by the ethics committee because this study was a retrospective experiment and did not involve patient privacy. All participating institutions were also informed and agreed the study.

Patient information

This retrospective study reviewed patient data collected from three hospitals: Hospital 1 (Shanghai Public Health Clinical Center), Hospital 2 (Shanghai Ruijin Hospital), and Hospital 3 (Yangzhi Rehabilitation Hospital). Data from Hospitals 1 and 2 were used to form the training and testing cohorts, while data from Hospital 3 served as the independent validation cohort. The sample sizes were as follows: 601 lesions for training, 151 for testing, and 76 for validation. In the training set, there were 393 category 0 [minimally invasive adenocarcinoma (MIA)], 127 category 1 [adenocarcinoma in situ (AIS)], and 156 category 2 [invasive adenocarcinoma (IAC)] samples. The test set included 99 category 0, 31 category 1, and 21 category 2 samples. The validation set comprised 51 category 0, 10 category 1, and 15 category 2 samples. Categorical variables were compared using the Chi-squared test, and continuous variables were compared using the Kruskal-Wallis test; the resulting P values are summarized in Table 1.

The inclusion criteria were: (I) routine chest CT obtained within one month prior to surgery or definitive diagnosis, with slice thickness ≤2 mm; and (II) if multiple nodules were present, each nodule was analyzed independently. The patient selection process and flowchart appear in Figure 1. Exclusion criteria included significant image artifacts on CT and use of intravenous contrast during acquisition. Relevant clinical information for each patient was integrated into the dataset for comprehensive analysis.

Figure 1 Flow chart of case sample selection. AIS, adenocarcinoma in situ; CT, computed tomography; IAC, invasive adenocarcinoma; MIA, minimally invasive adenocarcinoma.

CT image acquisition

All patients underwent unenhanced chest CT scanning across the four centers to ensure standardized lung imaging. At Hospital 1, images were acquired on two different scanners: a United-Imaging 760 CT (tube voltage 120 kVp; tube current modulation 42–126 mA; reconstructed slice thickness 1.0 mm) and a Siemens Emotion 16 CT (tube voltage 130 kVp; tube current modulation 34–123 mA; reconstructed slice thickness 1.0 mm). At Hospital 2, scans were obtained using a Philips iCT 256 (tube voltage 120 kVp; tube current modulation 161 mA; reconstructed slice thickness 1.0 mm) and a Philips Brilliance 16 CT (tube voltage 120 kVp; tube current modulation 219 mA; reconstructed slice thickness 1.0 mm). At Hospital 3, imaging was performed with a United-Imaging uCT960 + CT (tube voltage 140 kVp; tube current modulation 10–833 mA; reconstructed slice thickness 0.5 mm) and a Siemens Go Top CT (tube voltage ranging from 70 to 140 kVp in 10 kVp increments; tube current modulation 13–625 mA; reconstructed slice thickness 0.6 mm). All CT images were reconstructed using a standard lung-window kernel and de-identified prior to analysis.

Clinical features and key layer extraction

For three-dimensional segmentation of lung lesions, we employed a proprietary semi-automatic segmentation software. All segmentation outputs were initially reviewed by a radiologist with 6 years of experience and subsequently confirmed by a senior radiologist with twenty years of experience. Discrepancies between the two readers were resolved by consensus discussion. The software then automatically selected the central axial slice of each lesion and cropped a 64×64-pixel patch centered on the lesion (designated as IMG1).

Following segmentation, the same two radiologists independently evaluated the clinical radiographic features of each lesion, including margin definition (clear versus blurred), presence of lobulation and spiculation, and signs of pleural involvement (e.g., pleural tags or indentations) (25). The anatomical location of each nodule was recorded by lobe (left or right) and further classified into upper, middle, or lower lobe. Finally, a total of 944 radiomics descriptors—encompassing lesion size, shape metrics, first-order intensity statistics (histogram features), and higher-order texture features—were extracted from the region of interest (ROI) using the open-source PyRadiomics package (v. 3.0.1). Handcrafted radiomics features were used only as slice-wise descriptors to build a sequential representation and were not treated as standalone predictive inputs to the model.

Feature selection and integrated image dataset construction

Conventional CT-based lesion classification models often rely on a single static slice or a predefined subset of radiomics metrics, thereby neglecting the lesion’s evolving texture and morphology across adjacent slices. To address this limitation, we developed a spatiotemporal fusion framework. First, the central 64×64 region of each lesion (IMG1) is extracted to capture the most representative anatomical detail. From every contiguous slice, we compute 944 radiomics descriptors and apply variance filtering (variance <1) and inter-feature correlation filtering (|r|>0.90). Next, we employ minimum redundancy maximum relevance (mRMR) to select the 100 most informative descriptors, which are further refined via Lasso regression into a final set of 14 key features. These features are concatenated in slice order to form a dynamic feature sequence. The sequence is then processed by a BiLSTM-Attention network, explicitly modeling interslice textural evolution. For the first time in lesion classification, the network’s output vectors are projected onto two-dimensional feature maps using a Markov Random Field (MRF) approach. By fusing IMG1 with the resulting MRF-derived dynamic maps into a dual-channel 64×64 input, the architecture simultaneously learns spatial anatomy and temporal radiomics dynamics, thereby enhancing sensitivity to distinctions among MIA, AIS, and IAC.

Model training and evaluation

We implemented a dynamic radiomics fusion network (DRFN) that integrates anatomical and temporal radiomics information from the initial layers. The two 64×64 channels (the central CT slice and the MRF-mapped radiomics feature map) are concatenated and passed through three convolutional blocks (32→64→128 filters), each comprising Batch Normalization, ReLU activation, max-pooling, and Dropout, to learn rich cross-modal representations. After global average pooling, an attention gate adaptively balances anatomical versus radiomics cues, and the resulting feature vector is forwarded to a fully connected head (128→64→32→3) with Layer Normalization, ReLU, and Dropout, concluding with a softmax classifier. DRFN is trained using the Adam optimizer (learning rate =1×10⁻³, weight decay =1×10⁻⁵) together with a ReduceLROnPlateau scheduler and L₂ regularization to ensure rapid, stable convergence.

To validate the advantage of early dual-channel fusion, we compared DRFN against two baselines: (I) a standalone CNN that processes only the central slice and (II) a BiLSTM-Attention network that consumes the 14-feature radiomics sequences directly. Across cross-validation, DRFN consistently outperformed both baselines in accuracy and F1-score, demonstrating that joint spatiotemporal learning substantially enhances discriminative power. In addition, we applied class activation mapping (CAM) exclusively to DRFN: the resulting activation maps clearly highlighted both anatomical landmarks (e.g., spiculation, lobulation) and evolving radiomics patterns, underscoring the model’s superior interpretability and clinical relevance.

Results

Feature selection

All radiomics features were initially normalized to ensure comparability across samples. Features exhibiting near-zero variance were excluded, eliminating 127 redundant descriptors. A subsequent Spearman correlation analysis identified highly correlated descriptors. Those with absolute correlation coefficients greater than 0.90 were removed to reduce redundancy, leaving 237 candidate features. To further optimize feature relevance, the mRMR algorithm was applied, selecting the top 100 ranked features. Finally, LASSO regression with cross-validation was applied to the training cohort to identify the optimal subset of predictive features. The final selected feature set included various descriptors reflecting lesion morphology, intensity distribution, and texture patterns. Table 2 summarizes these selected features along with their corresponding descriptions and categories.

Model performance and comparison

Across all cohorts, DRFN achieved the strongest overall performance (Table 3; Figures 2,3).On the internal test cohort, DRFN reached an AUC of 0.974, accuracy 0.933, and F1-score 0.926, exceeding the BiLSTM-Attention model (AUC 0.883, F1 0.854) and the single-slice CNN baseline (AUC 0.771, F1 0.533).Performance differences became more apparent in the external validation cohort, where DRFN maintained an AUC of 0.925 and F1 of 0.878, whereas the BiLSTM-Attention model dropped to 0.827 (F1 0.861) and the CNN baseline declined substantially to 0.664 (F1 0.551). These results indicate that reliance solely on radiomics sequences (BiLSTM) or on static 2D images (CNN) limits cross-center generalizability, while the dual-domain fusion in DRFN better preserves discriminatory power.

Figure 2 Confusion matrices for model evaluation across training (A), testing (B), and validation (C) sets.

Figure 3 Multi-class ROC curves for model evaluation across training (A), testing (B), and validation (C) sets. AUC, area under the curve; ROC, receiver operating characteristic.

DRFN also demonstrated more balanced subtype performance. Class-wise metrics (Table 4) showed stable precision and recall across MIA, AIS, and IAC in the internal test cohort. In the external validation cohort, a moderate recall reduction for AIS (0.636) was observed, reflecting the limited sample size and inherent class imbalance, yet DRFN still outperformed the CNN baseline across all subtypes.

In addition to the overall AUC values, we further computed class-wise precision, recall, specificity, and F1-scores for MIA, AIS, and IAC across the training, internal test, and external validation cohorts (Table 4). The DRFN achieved balanced performance across all three subtypes on the internal test cohort. On the external validation cohort, recall for the minority AIS subtype showed a moderate decrease (0.636), reflecting the impact of limited sample size and class imbalance.

Interpretability analysis via CAM

Overall, CAM visualizations (Figure 4) confirm that DRFN reliably focuses on clinically salient features: on the anatomical channel, high activations consistently trace lesion margins, spiculation lobules, and pleural tags, while on the dynamic radiomics channel, attention peaks correspond to regions of pronounced textural change across slices (e.g., transitions from ground-glass to solid components). This dual emphasis demonstrates that DRFN genuinely integrates spatial morphology and temporal feature dynamics to pinpoint malignancy indicators.

Figure 4 Grad-CAM visualizations on dual-channel inputs for lung adenocarcinoma subtyping. Columns 1-3 correspond to MIA, AIS, and IAC cases, respectively. Panels (A-F) show correctly classified examples by DRFN, whereas panels (G-I) illustrate the CNN baseline’s Grad-CAM outputs—highlighting its failure to focus on lesion regions in the corresponding subtype.AIS, adenocarcinoma in situ; CNN, convolutional neural network; DRFN, dynamic radiomics fusion network; IAC, invasive adenocarcinoma; MIA, minimally invasive adenocarcinoma.

By contrast, when we apply the same CAM procedure to the single-channel CNN baseline (Figure 4G-4I), its attention frequently drifts toward adjacent vascular or pleural structures and underrepresents key lesion cores, especially in low-contrast nodules. These results highlight both the superior localization of DRFN and the limitations of static-image models—underscoring the value of our early dual-channel fusion approach.

Discussion

In this study, we aimed to approximate the radiologist’s slice-by-slice interpretive process by pairing the automatically selected central CT slice (IMG1) with a dynamically evolving sequence of radiomics signatures. The proposed DRFN demonstrates the effectiveness of this strategy. By integrating the 64×64 anatomical image with an MRF-derived dynamic feature map at the input stage, DRFN begins cross-modal representation learning from the very first convolutional layer. This early dual-domain fusion leads to substantially improved subtype discrimination: in the external validation cohort, DRFN achieved an AUC of 0.925 and an F1-score of 0.878, clearly outperforming the BiLSTM-Attention model (AUC 0.827, F1 0.861) and the static single-slice CNN (AUC 0.664, F1 0.551). The ability to jointly learn spatial structure and radiomics evolution not only enhances sensitivity and specificity but also markedly strengthens generalizability across multi-center datasets. Furthermore, when compared with representative studies in the literature that were evaluated on their respective independent datasets, our approach demonstrates consistently competitive—and often superior—performance, highlighting the value of dynamic feature integration in pulmonary adenocarcinoma subtyping.

Compared with earlier work on malignant lung diseases, our study offers several important methodological advances (26-28). Although fourteen Lasso-selected radiomics descriptors were used to construct the feature sequence, they function only as slice-wise quantitative summaries of lesion appearance. Collectively, these descriptors capture three major categories of information: (I) morphologic features that reflect lesion contour and margin complexity; (II) intensity-based measures that describe density distribution and attenuation changes; and (III) higher-order texture patterns that characterize structural heterogeneity. When arranged sequentially across axial slices, these features outline the evolution from ground-glass opacity to solid invasive components and provide a coherent representation of the transition from non-invasive to invasive adenocarcinoma. This organized radiomics progression forms the basis for generating the dynamic texture patterns visualized in our MRF maps. In this framework, handcrafted radiomics descriptors were used solely as slice-wise inputs to construct a multi-layer radiomics sequence, rather than as independent predictive features. The methodological novelty lies not in the radiomics descriptors themselves, but in the transformation of these slice-level signatures into a dynamic MRF-based feature map and its early fusion with the anatomical CT patch. This dual-domain fusion enables the network to jointly learn spatial structure and cross-slice progression, substantially enhancing subtype discrimination and generalizability across centers.

The MRF maps themselves offer an intuitive visualization of subtype-specific progression patterns MIA displays a remarkably consistent pattern with minimal layer-to-layer variation, AIS shows moderate edge and peritumoral fluctuations reflecting early invasive behavior, and IAC exhibits pronounced heterogeneity and rapid texture changes at the lesion margins, underscoring its aggressive nature. By converting temporal radiomics trends into spatially coherent textures, the MRF representation enables clinicians to appreciate lesion evolution within a single image, reinforcing the radiologist’s slice-by-slice diagnostic reasoning.

Critically, our CAM (29,30) interpretability analysis (Figure 4) reveals that, whereas the CNN baseline often directs attention to adjacent vessels or pleural lines, DRFN’s heatmaps consistently focus on lesion boundaries, spiculation lobules, and regions of maximal texture change across slices. In over 85% of correct cases, the top activation pixels overlap with known radiologic signs, and the dynamic channel consistently highlights slices where ground-glass first transitions to solid tissue. At the same time, CAM reveals residual weaknesses—in roughly 10% of challenging examples, attention diffuses to non-lesion structures and underweights hypodense cores—pointing the way toward refinements such as enhanced contrast normalization or specialized “core-focus” attention modules.

Although the DRFN achieved a perfect AUC of 1.000 on the training cohort, its performance on the internal test set (AUC =0.974) and external multi-center validation cohort (AUC =0.925) was slightly lower. This difference may reflect both the strong fitting ability of deep models in high-dimensional feature spaces and the inherent variability across multi-center imaging data. While the external validation results remain encouraging, we acknowledge this performance gap as a limitation and believe that expanding the size and diversity of multi-center cohorts will be important for further improving generalizability. Moreover, the overall cohort size is still relatively modest, particularly for rare or borderline phenotypes, which may constrain the full representation of subtype variability. The combined analysis of MRF-based feature maps and CAM-derived attribution maps also lacks a formal cross-modal quantitative framework. Occasional attention mis-localization in challenging lesions suggests that future work may benefit from more advanced normalization strategies, attention regularization, and larger, more balanced datasets. We anticipate that addressing these aspects will further strengthen the robustness of DRFN and support its eventual translation into clinical practice.

Conclusions

In this work, we developed a novel dual-channel fusion framework that combines central-slice anatomy with multilayer radiomics dynamic by mapping high-dimensional feature sequences into MRF-based spatial maps and integrating them at the network input. This early spatiotemporal fusion not only mirrors radiologists’ layer-by-layer assessment but also enables interpretable, high-fidelity classification of lung adenocarcinoma subtypes. The strong cross-center performance further highlights the potential of this dynamic fusion paradigm to support preoperative decision-making in clinical practice.

Acknowledgments

None.

Footnote

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1120/dss

Funding: This work was supported by the National Key Clinical Specialty Discipline Construction Project of China (No. Z155080000004), Shanghai Research Center of Rehabilitation Medicine (Top Priority Research Center of Shanghai) (No. 2023ZZ02027) and Shanghai Disabled Persons’ Federation Key Laboratory of Intelligent Rehabilitation Assistive Appliance and Technology.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1120/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the ethics committee of Shanghai University of Medicine and Health Sciences (No. 2019-ZYXM-04-321102****), and the written informed consent from patients was waived by the ethics committee because this study was a retrospective experiment and did not involve patient privacy. All participating institutions were also informed and agreed the study.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Christenson SA, Smith BM, Bafadhel M, Putcha N. Chronic obstructive pulmonary disease. Lancet 2022;399:2227-42. [Crossref] [PubMed]
2. Qian Y, Cai C, Sun M, Lv D, Zhao Y. Analyses of Factors Associated with Acute Exacerbations of Chronic Obstructive Pulmonary Disease: A Review. Int J Chron Obstruct Pulmon Dis 2023;18:2707-23. [Crossref] [PubMed]
3. Qi C, Sun SW, Xiong XZ. From COPD to Lung Cancer: Mechanisms Linking, Diagnosis, Treatment, and Prognosis. Int J Chron Obstruct Pulmon Dis 2022;17:2603-21. [Crossref] [PubMed]
4. Aredo JV, Luo SJ, Gardner RM, Sanyal N, Choi E, Hickey TP, et al. Tobacco Smoking and Risk of Second Primary Lung Cancer. J Thorac Oncol 2021;16:968-79. [Crossref] [PubMed]
5. Kurmi OP. Household Air Pollution, Passive Smoking, and Lung Cancer: Do We Know Enough about This Conundrum? Am J Respir Crit Care Med 2022;206:1065-7. [Crossref] [PubMed]
6. Berg CD, Schiller JH, Boffetta P, Cai J, Connolly C, Kerpel-Fronius A, Kitts AB, Lam DCL, Mohan A, Myers R, Suri T, Tammemagi MC, Yang D, Lam SInternational Association for the Study of Lung Cancer (IASLC) Early Detection and Screening Committee. Air Pollution and Lung Cancer: A Review by International Association for the Study of Lung Cancer Early Detection and Screening Committee. J Thorac Oncol 2023;18:1277-89. [Crossref] [PubMed]
7. Xue Y, Wang L, Zhang Y, Zhao Y, Liu Y. Air pollution: A culprit of lung cancer. J Hazard Mater 2022;434:128937. [Crossref] [PubMed]
8. Mederos N, Friedlaender A, Peters S, Addeo A. Gender-specific aspects of epidemiology, molecular genetics and outcome: lung cancer. ESMO Open 2020;5:e000796. [Crossref] [PubMed]
9. Vicidomini G. Current Challenges and Future Advances in Lung Cancer: Genetics, Instrumental Diagnosis and Treatment. Cancers (Basel) 2023;15:3710. [Crossref] [PubMed]
10. Lam S, Bai C, Baldwin DR, Chen Y, Connolly C, de Koning H, Heuvelmans MA, Hu P, Kazerooni EA, Lancaster HL, Langs G, McWilliams A, Osarogiagbon RU, Oudkerk M, Peters M, Robbins HA, Sahar L, Smith RA, Triphuridet N, Field J. Current and Future Perspectives on CT Screening for Lung Cancer: A Road Map for 2023-2027 from the IASLC. J Thorac Oncol 2024;19:36-51. [Crossref] [PubMed]
11. Hata A, Hino T, Yanagawa M, Nishino M, Hida T, Hunninghake GM, Tomiyama N, Christiani DC, Hatabu H. Interstitial Lung Abnormalities at CT: Subtypes, Clinical Significance, and Associations with Lung Cancer. Radiographics 2022;42:1925-39. [Crossref] [PubMed]
12. Prosper AE, Kammer MN, Maldonado F, Aberle DR, Hsu W. Expanding Role of Advanced Image Analysis in CT-detected Indeterminate Pulmonary Nodules and Early Lung Cancer Characterization. Radiology 2023;309:e222904. [Crossref] [PubMed]
13. Hussain S, Mubeen I, Ullah N, Shah SSUD, Khan BA, Zahoor M, Ullah R, Khan FA, Sultan MA. Modern Diagnostic Imaging Technique Applications and Risk Factors in the Medical Field: A Review. Biomed Res Int 2022;2022:5164970. [Crossref] [PubMed]
14. Nooreldeen R, Bach H. Current and Future Development in Lung Cancer Diagnosis. Int J Mol Sci 2021;22:8661. [Crossref] [PubMed]
15. Nazir A, Hussain A, Singh M, Assad A. A novel approach in cancer diagnosis: integrating holography microscopic medical imaging and deep learning techniques-challenges and future trends. Biomed Phys Eng Express 2025;11: [Crossref] [PubMed]
16. Jia H, Zhang J, Ma K, Qiao X, Ren L, Shi X. Application of convolutional neural networks in medical images: a bibliometric analysis. Quant Imaging Med Surg 2024;14:3501-18. [Crossref] [PubMed]
17. Kaifi R. Enhancing brain tumor detection: a novel CNN approach with advanced activation functions for accurate medical imaging analysis. Front Oncol 2024;14:1437185. [Crossref] [PubMed]
18. Wang H, Zhu H, Ding L. Accurate classification of lung nodules on CT images using the TransUnet. Front Public Health 2022;10:1060798. [Crossref] [PubMed]
19. Yu JD, Li TT, Shi XR, Zhao ZY, Chen M, Yu Zhang, Wang JY, Yao ZJ, Fang L, Hu B. ETMO-NAS: An efficient two-step multimodal one-shot NAS for lung nodules classification. Biomed Signal Process Control 2025;104:107479.
20. Lin CY, Guo SM, Lien JJ, Lin WT, Liu YS, Lai CH, Hsu IL, Chang CC, Tseng YL. Combined model integrating deep learning, radiomics, and clinical data to classify lung nodules at chest CT. Radiol Med 2024;129:56-69. [Crossref] [PubMed]
21. Tas MO, Yavuz HS. Enhancing Lung Cancer Survival Prediction: 3D CNN Analysis of CT Images Using Novel GTV1-SliceNum Feature and PEN-BCE Loss Function. Diagnostics (Basel) 2024;14:1309. [Crossref] [PubMed]
22. Neal Joshua ES, Bhattacharyya D, Chakkravarthy M, Byun YC. 3D CNN with Visual Insights for Early Detection of Lung Cancer Using Gradient-Weighted Class Activation. J Healthc Eng 2021;2021:6695518. [Crossref] [PubMed]
23. Han D, Chen Y, Li X, Li W, Zhang X, He T, Yu Y, Dou Y, Duan H, Yu N. Development and validation of a 3D-convolutional neural network model based on chest CT for differentiating active pulmonary tuberculosis from community-acquired pneumonia. Radiol Med 2023;128:68-80. [Crossref] [PubMed]
24. Mikhael PG, Wohlwend J, Yala A, Karstens L, Xiang J, Takigami AK, Bourgouin PP, Chan P, Mrah S, Amayri W, Juan YH, Yang CT, Wan YL, Lin G, Sequist LV, Fintelmann FJ, Barzilay R. Sybil: A Validated Deep Learning Model to Predict Future Lung Cancer Risk From a Single Low-Dose Chest Computed Tomography. J Clin Oncol 2023;41:2191-200. [Crossref] [PubMed]
25. van Griethuysen JJM, Fedorov A, Parmar C, Hosny A, Aucoin N, Narayan V, Beets-Tan RGH, Fillion-Robin JC, Pieper S, Aerts HJWL. Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Res 2017;77:e104-7. [Crossref] [PubMed]
26. Huang B, Sollee J, Luo YH, Reddy A, Zhong Z, Wu J, Mammarappallil J, Healey T, Cheng G, Azzoli C, Korogodsky D, Zhang P, Feng X, Li J, Yang L, Jiao Z, Bai HX. Prediction of lung malignancy progression and survival with machine learning based on pre-treatment FDG-PET/CT. EBioMedicine 2022;82:104127. [Crossref] [PubMed]
27. Ma Y, Li J, Xu X, Zhang Y, Lin Y. The CT delta-radiomics based machine learning approach in evaluating multiple primary lung adenocarcinoma. BMC Cancer 2022;22:949. [Crossref] [PubMed]
28. Zhang Y, Liu H, Chang C, Yin Y, Wang R. Machine learning for differentiating lung squamous cell cancer from adenocarcinoma using Clinical-Metabolic characteristics and 18F-FDG PET/CT radiomics. PLoS One 2024;19:e0300170. [Crossref] [PubMed]
29. Iqbal S, Qureshi AN, Alhussein M, Aurangzeb K, Anwar MS AD-CAM. Enhancing Interpretability of Convolutional Neural Networks with a Lightweight Framework - From Black Box to Glass Box. IEEE J Biomed Health Inform 2023; Epub ahead of print. [Crossref]
30. Pion E, Karnosky J, Boscheck S, Wagner BJ, Schmidt KM, Brunner SM, Schlitt HJ, Aung T, Hackl C, Haerteis S 3D. In Vivo Models for Translational Research on Pancreatic Cancer: The Chorioallantoic Membrane (CAM) Model. Cancers (Basel) 2022;14:3733. [Crossref] [PubMed]