To investigate the prognostic value of baseline MRI characteristics based on the DISTANCED structured report in patients with MRI-defined T3 (mrT3) rectal cancer. We retrospectively analyzed 190 patients with mrT3 rectal cancer who received neoadjuvant therapy followed by surgical resection between December 2014 and June 2023, with a median follow-up of 43 months (range: 2–126 months). Baseline MRI features, clinical characteristics, and follow-up data were assessed. The patients were randomly assigned to a training cohort (n = 133) or a validation cohort (n = 57) in a 7:3 ratio. Cox regression analyses were performed to identify independent risk factors for disease-free survival (DFS), which were then used to construct a nomogram in the training cohort. The nomogram was independently validated using the validation cohort. Harrell’s concordance index (C-index) and time-independent receiver operating characteristic (ROC) analysis were used to evaluate the model’s discrimination. For patient stratification, the DFS rates of high- and low-risk patients were calculated using the Kaplan-Meier method. Baseline MRI-determined lateral lymph node (mrLLN) metastasis, circumference resection margin (mrCRM) involvement, and tumor deposits (mrTDs) were independent predictors of DFS. The nomogram exhibited good discrimination, with a C-index of 0.805 (95
To investigate the non-invasive predictive value of quantitative spectral parameters from photon-counting computed tomography (PCCT) for epidermal growth factor receptor (EGFR) mutation status and its predominant subtypes (19Del and L858R) in patients with lung adenocarcinoma. A total of 72 patients with pathologically confirmed lung adenocarcinoma who underwent pretreatment PCCT were retrospectively enrolled. Arterial and venous CT attenuation values at 40 keV, 70 keV, and 100 keV (A/V-40 keV, A/V-70 keV, A/V-100 keV) were measured on virtual monoenergetic images, while arterial/venous iodine concentration (IC) and dual-energy index (DEI) of lesions were measured on iodine maps and spectral post-processing (SPP) images, respectively. Normalized iodine concentration (NIC) and spectral curve slope (λHU) were further calculated. Receiver operating characteristic (ROC) curve analysis and binary logistic regression were performed to evaluate the predictive performance and independent predictive value of PCCT parameters for discriminating EGFR-mutant vs. wild-type tumors, as well as 19Del vs. L858R subtypes. Of 72 patients, 37 (51.4
To construct an integrated prognostic model combining intra‑ and peritumoral radiomics with MRI features for predicting progression‑free survival (PFS) in soft tissue sarcomas (STS). This retrospective study included 305 patients with trunk or extremity STS. Radiomic features were extracted from intratumoral and peritumoral regions at expansion distances of 3 mm, 5 mm, and 10 mm on both contrast‑enhanced T1‑weighted (CE‑T1WI) and fat‑suppressed T2‑weighted (FS‑T2WI) sequences. Features were selected via ICC, univariate Cox, and LASSO Cox regression. Prognostic models were constructed by integrating Rad‑scores with significant MRI features, and evaluated using C‑index, calibration curves, and decision curve analysis. Subgroup analyses were performed stratified by imaging phenotype (cellular vs. non‑cellular). Correlation analyses between peritumoral Rad‑scores and semantic MRI features were also conducted. The 3‑mm peritumoral expansion yielded the highest predictive performance among the three distances. Integrated intratumoral–peritumoral (3 mm) models from both sequences outperformed single‑region models. The 3‑mm margin consistently demonstrated superior performance in both cellular and non‑cellular phenotypes. The final multimodal model, combining dual‑sequence Rad‑scores with peritumoral enhancement, achieved a C‑index of 0.888 in the validation set. Correlation analyses reveal that peritumoral Rad‑scores are most strongly associated with peritumoral T2 hyperintensity, with differential correlation patterns between CE‑T1WI and FS‑T2WI supporting their complementary roles. We developed a robust multimodal radiomics model integrating intra‑ and peritumoral (3 mm) features with key MRI characteristics. The optimality of the 3‑mm margin was validated across major STS phenotypes. This non‑invasive tool facilitates personalized risk stratification and may inform adjuvant treatment decisions, supporting clinical translation in STS management.
This study aimed to demonstrate tumor immune microenvironment (TIME) subtypes as prognostic factors for local recurrence-free survival (LRFS) in esophageal squamous cell carcinoma (ESCC) patients treated with definitive chemoradiotherapy (dCRT) and to develop an interpretable, TIME-guided subregional radiomics model for predicting LRFS. 986 ESCC patients from three centers were enrolled. For a subset with available biopsies, multiplex immunofluorescence was used to assess CD3, CD8, and PD-L1 expression, defining prognostic TIME subtypes. Pre-treatment contrast-enhanced CT (CECT) images were used for tumor segmentation, intratumoral subregion partitioning and radiomic feature extraction. Features associated with TIME subtypes were selected to build a subregional radiomic signature (Subrad signature) for LRFS prediction using an advanced machine learning framework. This signature was integrated with clinical factors to construct a Clinic-Subregional radiomic model (CliSubrad-M). A “PD-L1-low/TILs-high” TIME subtype was identified as an independent favorable prognostic factor for LRFS (HR = 0.34, P < 0.001). The Subrad signature, comprising seven TIME-related features from various subregions, demonstrated robust predictive performance (C-indexes: 0.631–0.637). The integrated CliSubrad-M (built with Random Survival Forest) achieved the best performance, with C-indexes of 0.702 (training), 0.679 (internal testing), 0.653 (external testing1), and 0.625 (external testing2), significantly outperforming the clinical model (all P < 0.05). The model effectively stratified patients into distinct risk groups with significant differences in LRFS. SHAP analysis confirmed the signature’s strong contribution and enhanced model interpretability. The developed interpretable subregional radiomics model, integrating TIME biology with clinical factors, provides a non-invasive tool for predicting LRFS in ESCC received dCRT.
Standard [¹⁸F]FDG PET/CT has suboptimal sensitivity for assessing viable brain metastases due to high physiological brain background. Fibroblast activation protein (FAP)-targeted PET, with minimal brain expression, may overcome this limitation. This study aimed to compare the diagnostic performance of [⁶⁸Ga]Ga-FAPI-2286 versus [¹⁸F]FDG PET/CT for detecting viable brain metastases in lung cancer and evaluate their complementary value. In this prospective intra-individual study, 24 lung cancer patients with brain metastases underwent both [⁶⁸Ga]Ga-FAPI-2286 and [¹⁸F]FDG PET/CT within 3 days. Images were independently and blindly reviewed by two senior nuclear medicine physicians. Diagnostic performance (sensitivity, specificity, accuracy) was evaluated per-lesion and per-patient against a composite reference standard (characteristic imaging features, ≥ 3-month follow-up, comprehensive clinical evaluation). Quantitative parameters (SUVmax, target-to-background ratio [TBR], background uptake [SUVbgd]) and complementary detection patterns were analyzed. Among 43 lesions (36 viable, 7 non-viable), [⁶⁸Ga]Ga-FAPI-2286 demonstrated significantly higher per-lesion sensitivity (83.3
Neoadjuvant immunotherapy plus chemoradiotherapy (nICRT) for esophageal squamous cell carcinoma (ESCC) received increasing attention because of high pathological complete response (pCR) rate. Accurately assessing residual disease after nICRT is beneficial for dynamic monitoring of patients. This exploratory, single-arm, phase II clinical study aimed to evaluate the diagnostic efficacy of 18F-FDG PET/CT for achieving pathological complete response (pCR) after nICRT for ESCC, and its prognostic value for survival. From January 12, 2021 to February 21, 2023, a total of 21 resectable ESCC patients receiving nICRT were enrolled and underwent baseline PET/CT scans (scan-1) and surgery. 19 patients underwent scans (scan-2) before surgery. The recorded PET/CT parameters included maximum, mean, and peak standardized uptake values (SUVmax, SUVmean, and SUVpeak), metabolic tumor volume (MTV), and total lesion glycolysis (TLG). The diagnostic performance of PET/CT parameters was analyzed between the patients with pCR of the primary tumor (pCR-tu) and non-pCR-tu. The Cox regression and Kaplan-Meier method were performed for multivariable analysis and survival analyses, respectively. 13(61.9
Photon-counting CT (PCCT) is an emerging imaging modality that offers improved spatial resolution, soft tissue contrast, material decomposition, and dose efficiency over conventional CT with energy-integrating detectors (EID-CT). For patients with abdominal malignancies, who routinely undergo cross-sectional imaging for diagnosis, staging, response assessment, and surveillance, PCCT has the potential to enhance the detection and characterization of contrast-enhanced tumors and improve treatment response assessment. The objectives of this review are to provide a comprehensive overview of the technical principles of PCCT, optimization strategies, advantages in abdominal oncologic imaging, and current clinical applications geared towards clinical radiologists. We also address implementation challenges and discuss future directions for integrating PCCT into routine oncologic care. As PCCT utilization increases, its impact on abdominal cancer diagnosis and management is expected to grow, offering new opportunities for precision medicine and personalized imaging strategies. Not applicable.
To evaluate whether spatial topology features (STFs) from baseline 18 F-FDG PET/CT provide prognostic value beyond conventional burden and dissemination metrics in disseminated diffuse large B-cell lymphoma (DLBCL). This multicenter study included 243 patients (training, n = 161; validation, n = 82) with de novo DLBCL (≥ 4 lesions). An AI-assisted, human-in-the-loop pipeline was used to automatically detect lesions and construct 3D spatial networks. Six hierarchical models predicted 2-year progression-free survival, comparing a standard burden model (M3), a topology model (M4), and a fusion model (M5). Performance was assessed via AUC, Brier score, and calibration slope. Clinical utility was analyzed using Decision Curve Analysis (DCA), and biological correlations were tested against molecular subtypes. In the external validation cohort, the clinical baseline model (M0) showed limited efficacy (AUC 0.620). M4 showed similar discrimination to M3 (AUC 0.819 vs. 0.828; p = 0.68), but better specificity (76.4
To evaluate the adjunctive value of 18F-FDG PET/CT for assessing residual metastatic lymph nodes (MLN) before initial radioiodine therapy in postoperative DTC patients with high clinical suspicion of residual nodal disease. This retrospective study included 835 postoperative patients with DTC who underwent 18F-FDG PET/CT before initial 131I therapy because of elevated stimulated thyroglobulin (sTg) and/or suspicious conventional neck imaging. MLN status was determined using a composite reference standard. Patient-level logistic regression was used to evaluate associated factors, and receiver operating characteristic (ROC) analysis was performed to assess discrimination and determine optimal cutoffs. A sensitivity analysis reassessed diagnostic performance using a reference assessment that excluded PET/CT information. Of the 835 patients, 452 (54.1
This study aimed to investigate the value of the extracellular volume fraction (ECV) derived from dual-energy computed tomography (DECT) in predicting the chemotherapeutic response in patients with pancreatic ductal adenocarcinoma (PDAC). This study included 112 patients with pathologically confirmed PDAC who underwent dynamic contrast-enhanced DECT prior to chemotherapy. Iodine concentrations were measured on delayed-phase iodine maps within the tumor and aorta, and the ECV was calculated on the basis of the iodine concentration. Treatment response to chemotherapy was assessed according to the RECIST 1.1 criteria. Univariate and multivariate analyses were conducted to identify predictors of treatment response. Receiver operating characteristic (ROC) curves were used to evaluate the predictive performance of the identified variables. The tumor ECV was significantly lower in the response group (23.64
The aim of this research is to develop and evaluate a novel feature selection method based on a local search algorithm (LSFS) in combination with a genetic algorithm (GA) for cancer classification using high-dimensional gene expression data. The study spanned a period between 2022 and 2024, using data from two medical research centers (Sofia, Bulgaria, and Moscow, Russia). More than 30,000 patient records were analyzed using six open datasets as source data. These datasets included information on blood, breast, colon, and other types of cancer from the GEO and TCGA databases. To prepare the data for analysis, the study employed techniques such as normalization, outlier elimination, and logarithmic transformation. Modern machine learning algorithms, such as random forests, support vector machines, and neural networks, were utilized to assess the effectiveness of the proposed method. The LSFS + GA method demonstrated the ability to maximize accuracy, sensitivity, and specificity, as well as other classification metrics. Specifically, classification accuracy improved by 8–15
Accurate assessment of Ki-67 expression levels in breast cancer is crucial for determining prognosis and making informed treatment decisions. Current immunohistochemical methods relying on needle biopsy introduce sampling errors due to tumor spatial heterogeneity, making the development of non-invasive, precise preoperative prediction methods of significant clinical importance. This study aims to explore and compare advanced deep learning models based on dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) for noninvasive assessment of Ki-67 expression. This retrospective study analyzed preoperative DCE-MRI data from 308 patients with histologically confirmed breast cancer. Adjacent slices centered on the tumor’s most significant cross-section were obtained to create a 2.5-dimensional (2·5D) dataset. We innovatively developed two deep learning models using the same dataset (1): a Multi-Instance Learning (MIL) model that combines slice-level predictive features with Predictive Likelihood Histogram (PLH) and Bag-of-Words (BoW) techniques (2); a Transformer-based fusion model that directly captures global contextual relationships between slices via self-attention mechanisms. The predictive performance of both models was systematically compared with traditional radiomics and clinical models. On an independent test set, the Transformer fusion model demonstrated optimal predictive performance with an area under the curve (AUC) of 0.875, achieving accuracy, sensitivity, and specificity of 0.839, 0.848, and 0.833, respectively. The MIL model ranked second (AUC = 0.825), with both models significantly outperforming traditional radiomics models (AUC = 0.698) and clinical models (AUC = 0.648). Deep learning models based on 2·5D DCE-MRI, especially Transformer models that achieve global feature fusion through self-attention mechanisms, can effectively and non-invasively predict Ki-67 expression status in breast cancer, surpassing traditional methods. This model shows potential as a reliable tool to help clinicians accurately assess tumor proliferation activity before surgery.
Abstract Purpose Portal hypertension-related parameters are known to have important prognostic implications in patients with hepatocellular carcinoma (HCC). Ascites has consistently been associated with impaired outcomes. However, the prognostic relevance of quantified ascites volume in patients undergoing Atezolizumab/Bevacizumab remains unclear. Methods 90 patients with HCC treated with Atezolizumab/Bevacizumab between 2020 and 2024 were retrospectively included. Ascites was assessed on baseline contrast-enhanced CT imaging using AI-based pre-segmentation followed by manual correction to quantify ascites volume. Overall survival (OS) was analyzed using Kaplan–Meier analysis and Cox regression. Ascites volume was analyzed both as a continuous variable and using predefined volumetric cutoffs (< 613.5 mL and ≥ 613.5 mL). Results Ascites was present in 39% of patients, with a median ascites volume of 344 mL (IQR 29–929). When analyzed as a continuous variable, ascites volume was significantly associated with impaired OS in univariate analysis (Hazard Ratio (HR) 2.88, p < 0.001) but not after multivariate adjustment ( p = 0.214). Stratification by ascites volume revealed a stepwise reduction in OS, with patients without ascites showing the best survival, followed by patients with low-volume ascites (< 613.5 ml), and patients with high-volume ascites (≥ 613.5 ml) demonstrating the poorest outcomes (log-rank p < 0.001). In multivariate analysis, high-volume ascites remained independently associated with impaired OS (HR 4.01, p < 0.001), whereas low-volume ascites did not. Conclusion Ascites volume showed a threshold-dependent prognostic effect in patients with HCC treated with Atezolizumab/Bevacizumab. Volumetric staging identifies high-volume ascites as an independent marker of poor outcome and may provide more clinically relevant risk stratification than binary ascites assessment. Clinical trial number Not applicable.
To develop and validate a machine learning model integrating habitat imaging features before and after neoadjuvant chemoimmunotherapy (NACI) to predict pathological complete response (pCR) to NACI in head and neck squamous cell carcinoma (HNSCC). A multicenter retrospective study was conducted involving 520 HNSCC patients who received NACI across three hospitals. Unlike conventional Delta-Radiomics, we developed Delta-Habitat model that integrates Pre-NACI and Post-NACI MRI spatial habitat radiomics features, and dynamically learns task-specific contributions from each timepoint. Predictive performance was assessed using the area under the receiver operating characteristic curve (AUC). SHapley Additive exPlanations (SHAP) was used to quantify the contribution of individual features to model predictions. We assessed and compared the diagnostic performance of senior and junior radiologists, both with and without the aid of the Delta-Habitat model. Delta-Habitat model outperformed Pre-NACI feature-based models and traditional radiomics-based Delta-Rad model, with AUC of 0.828, 0.825 and 0.813 in validation and two external test cohorts respectively. Post-treatment neutrophil-to-lymphocyte ratio (Post-NLR) was identified as an independent predictor of pCR. Integrating Delta-Radiomics and clinical features into Delta-Habitat model did not yield a significant improvement in predictive performance, the combined model showed slightly improved AUC over Delta-Habitat alone, reaching 0.877 (train), 0.834 (val), 0.837 (test1), and 0.834 (test2). SHAP analysis revealed that Post-NACI features contributed more substantially to model predictions than Pre-NACI features. The Delta-Habitat model consistently exhibited better performance than radiologists in comparative evaluations. Moreover, senior and junior radiologists both showed increases in AUC values with the assistance of the Delta-Habitat model, by 0.073 and 0.177 respectively on the internal validation cohort and by 0.087 and 0.142 on the external test cohort. The longitudinal MRI-based habitat model can provide valuable information for predicting the response to NACI in HNSCC and enhance radiologists’ diagnostic performance. • The Delta-Habitat model, incorporating Pre-NACI and Post-NACI habitat imaging features, can predict the pCR in HNSCC patients. • The developed model can assist radiologists to improve the interpretation of pCR in HNSCC patients.
To develop a short-time Fourier transform (STFT)-based perfusion fingerprinting method combined with spectral-habitat analysis for characterizing pixel-wise perfusion heterogeneity on contrast-enhanced ultrasound (CEUS) and to evaluate its ability to differentiate benign and malignant breast lesions. Pixel-wise CEUS time-series data were transformed into the time-frequency domain using STFT. Principal component analysis (PCA) was then applied to reduce the dimensionality of STFT-derived features and generate perfusion fingerprinting maps that preserved spatial and temporal perfusion information. Spectral-habitat maps were generated using K-means clustering to identify perfusion-related tumor subregions. A triple-branch multi-layer fusion model was constructed to integrate full-region and habitat-derived features for lesion classification. To provide an exploratory conventional CEUS quantitative baseline, TIC-derived perfusion features were additionally extracted and evaluated using machine learning classifiers. Model performance was evaluated using receiver operating characteristic (ROC) analysis, decision curve analysis (DCA), and confusion matrix. The proposed STFT-based perfusion fingerprinting method enabled a more comprehensive characterization of tumor perfusion heterogeneity by preserving pixel-wise spatiotemporal dynamics of CEUS signals. CEUS-derived spectral-habitat maps provided visual representations of intratumoral perfusion heterogeneity and improved the interpretability of the model outputs. In this cohort, the triple-branch multi-layer fusion model achieved an AUC of 0.953, with accuracy of 0.909, precision of 0.907, recall of 0.925, and F1 score of 0.912. Exploratory TIC-based conventional CEUS models achieved AUCs ranging from 0.7677 to 0.8186. DCA suggested potential clinical utility of the proposed model across a range of threshold probabilities, and confusion matrix analysis showed favorable classification performance. The STFT-based perfusion fingerprinting method combined with spectral-habitat analysis provides a promising CEUS-based framework for characterizing perfusion heterogeneity in breast lesions. By preserving pixel-wise spatiotemporal information, this approach may provide complementary quantitative information for more interpretable differentiation between benign and malignant lesions.
To compare the diagnostic performance and complication profiles of CT-guided core-needle biopsy (CNB) and fine-needle aspiration (FNA) for pulmonary ground-glass nodules (GGNs). Consecutive patients who underwent CT-guided percutaneous biopsy of pulmonary GGNs between January 2021 and August 2022 were screened. Of 657 patients assessed, 538 were included (CNB, n = 448; FNA, n = 90). Six FNA results classified as atypical cells were considered indeterminate and excluded from binary diagnostic-performance analyses, leaving 532 classifiable cases. Final diagnoses were established by surgical pathology or imaging follow-up of at least 12 months. Diagnostic performance was compared without adjustment, whereas complication outcomes were assessed using multivariable logistic regression and inverse probability of treatment weighting (IPTW). Among the 532 classifiable biopsies, the overall accuracy, sensitivity, and specificity were 93.2
To evaluate the diagnostic value of clinical and intravoxel incoherent motion (IVIM) parameters for differentiating large anterior mediastinal invasive masses (≥ 3 cm), particularly thymic epithelial tumors (TETs) and lymphomas. This prospective study included 92 patients (mean age, 44.8 ± 15.5 years) with pathologically confirmed TETs (n = 52) or lymphomas (n = 40) who underwent routine MRI, diffusion-weighted imaging, and IVIM sequences. Quantitative parameters, including the apparent diffusion coefficient (ADC), true diffusion coefficient (D), pseudo-diffusion coefficient (D*), and perfusion fraction (f), were measured. Univariate and multivariate logistic regression analyses were used to construct predictive models. Diagnostic performance was evaluated using receiver operating characteristic (ROC) analysis, with area under the curve (AUC) compared by the DeLong test. Internal validation was performed using bootstrap resampling, and multicollinearity was assessed using variance inflation factors. Compared with TET, lymphoma was associated with younger age, higher female proportion, higher LDH, larger volume, more frequent pleural and/or lung invasion, and higher D * values (all p < 0.05). Age, sex, LDH level, lung and/or pleural infiltration, and f in the IVIM model were risk predictive factors for distinguishing lymphoma from TET. Among all models, the clinical MRI-IVM model achieved the highest diagnostic performance numerically (AUC = 0.946, 95
We aimed to evaluate the prognostic value of baseline 18F-FDG PET/CT metabolic parameters and to construct a risk stratification model for recurrent/metastatic nasopharyngeal carcinoma (rmNPC) patients receiving first-line PD-1 inhibitors plus chemotherapy. A total of 148 rmNPC patients from two centers were included. Whole-body total lesion glycolysis (wTLG), whole-body metabolic tumor volume (wMTV), and maximal standardized uptake values (SUVmax) were assessed. Predictive performance was evaluated via receiver operating characteristic (ROC) curve, while progression-free survival (PFS) and overall survival (OS) were identified using Kaplan-Meier and Cox regression analyses. A recursive partitioning analysis (RPA) model was subsequently developed and validated for risk stratification. Lower baseline wTLG and wMTV were associated with better response and longer PFS and OS, with wTLG showing the highest predictive power (AUC = 0.776). The RPA model stratified patients into low-risk (low wTLG), medium-risk (high wTLG without liver metastasis), and high-risk (high wTLG with liver metastasis) groups, demonstrating significantly different median PFS (30.7, 15.9, 7.1 months) and OS (43.7, 29.9, 14.7 months; all p < 0.001), and achieving an AUC of 0.839 for predicting 3-year disease progression. Radiotherapy improved survival outcomes in the low-risk group but not in the high-risk group. Analysis of 213 individual lesions revealed heterogeneous responses, with lymph node metastases showing the best response and liver metastases the poorest. Baseline PET/CT metabolic parameters, particularly wTLG, serve as powerful predictors of survival in NPC patients receiving chemoimmunotherapy. The wTLG-based RPA model provides a robust tool for risk stratification and holds promise for guiding individualized clinical decision-making.
Clinicians need a reliable, noninvasive tool that can predict the risk of gastrointestinal stromal tumor (GIST) recurrence preoperatively. We aimed to develop a machine learning model based on preoperative contrast-enhanced CT (CECT) to predict recurrence-free survival (RFS) in GIST patients who underwent radical resection. A total of 192 patients with intermediate- and high-risk GISTs who underwent radical resection and subsequently received adjuvant imatinib were included, with a minimum follow-up duration of 24 months. A deep learning model (Modelradiomic) based on preoperative CECT was built to predict RFS, which was compared with the Armed Forces Institute of Pathology (AFIP) index (ModelAFIP). The C-index and time-dependent receiver operating characteristic curves were estimated via the Kaplan-Meier method and compared via the log-rank test. The C-index values of the Modelradiomic in the training, testing and validation cohorts ranged from 0.678 to 0.763, whereas the areas under the curves (AUCs) at 3 and 5 years were 0.744–0.790 and 0.703–0.833, respectively. The C-index values of ModelAFIP in the above datasets were 0.539–0.660, and the AUCs at 3 and 5 years were 0.497–0.649 and 0.503–0.630. Modelradiomic demonstrated a higher C-index than ModelAFIP did in the training and external validation cohorts (p = 0.010 and 0.042, respectively). Modelradiomic presented a greater AUC value than ModelAFIP did at the 5th year in the training cohort (p = 0.009). We found the machine learning-based preoperative CECT performed better than the AFIP index in prediction of RFS of GIST patients, especially at the 5th year, predicting recurrence risk in patients who underwent radical resection and receiving adjuvant therapy. This model may serve as a non-invasive tool to identify high-risk individuals who require more intensive surveillance and personalized management following radical resection.
To develop an intratumor heterogeneity (ITH)-aware deep learning (DL) framework for preoperative prediction of lymph node metastasis (LNM) in laryngeal squamous cell carcinoma (LSCC). This multi-center retrospective study, which included 381 patients, proposed a ITH-aware DL framework integrated with adaptive habitat mapping to automatically decode spatial ITH from primary tumor contrast-enhanced CT images.The predictive performance of the ITH-aware DL model was compared against conventional radiomics and 3D DL model. A stacking ensemble model combining these three methods was also constructed. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), calibration curves, and decision curve analysis (DCA). The prognostic value for overall survival (OS) stratification was evaluated. The ITH-aware DL model achieved superior preoperative LNM prediction in external testing (AUC: 0.745–0.766), outperforming both radiomics (AUC: 0.682–0.694) and 3D DL models (AUC: 0.718–0.732). The stacking ensemble integrating these models attained the highest diagnostic accuracy (AUC: 0.786–0.804). Critically, both the ITH-aware DL and stacking models effectively stratified patients into distinct risk groups with significantly different OS outcomes. The proposed ITH-aware DL framework provides a robust, automated tool for preoperative LNM prediction in LSCC. By enabling accurate risk stratification, it holds significant potential to inform personalized surgical planning and improve treatment strategies.
