Colonic Neoplasm
Conditions
Keywords
Colon cancer, deep learning, CT scan, TNM classification, radiology, AI, machine learning, tumor staging, colorectal neoplasms
Brief summary
This study aims to improve the classification of colon tumors using deep learning models trained on CT scans, specifically to distinguish between T1-T2 vs. T3-T4 stages and N- vs. N+ lymph node involvement. This classification is critical to guide preoperative treatment such as chemotherapy or immunotherapy. Given the limited accuracy of radiologists in current staging practice, automated image-based AI tools could enhance diagnostic precision and reproducibility, leading to more personalized and effective treatment planning. The investigator will develop and validate convolutional and transformer-based deep learning models using a large annotated dataset from multiple centers. Secondary objectives include fine-grained staging (T1 to T4), subgroup-specific models (MSS vs MSI), and predictive models for surgical
Detailed description
This is a retrospective, non-interventional, observational study evaluating the use of deep learning methods to improve preoperative CT-based TNM staging in patients with colon cancer. The study is conducted across multiple sites within the AP-HP hospital network (Paris, France) and uses data extracted from the institutional Health Data Warehouse. Radiologic accuracy in assessing tumor stage (T) and lymph node status (N) remains limited, despite being critical for selecting neoadjuvant treatments. Artificial intelligence models trained on annotated imaging data may provide more consistent, reproducible, and accurate classification. The study cohort includes adult patients who underwent colon resection between January 2017 and November 2024, with a preoperative CT scan and corresponding pathology report. Eligible cases are identified using standardized diagnostic (ICD-10) and procedural (CCAM) codes. Imaging and clinical data are de-identified prior to analysis. Several AI model architectures will be tested, including 3D convolutional neural networks and transformer-based approaches. CT scans will be pre-processed using standard pipelines; pathology labels will be extracted using natural language processing (NLP) techniques or manual review when needed. Model performance will be assessed through cross-validation and evaluated using AUC, F1-score, sensitivity, and specificity. Exploratory analyses will include fine-grained tumor staging and the potential prognostic value of image-based features for clinical outcomes such as survival. No study-related procedures are performed. All analyses are conducted on existing data, in compliance with French data protection and ethical regulations.
Interventions
None listed
Sponsors
Assistance Publique - Hôpitaux de Paris
Institut National de Recherche en Informatique et en Automatique
Study design
Observational model
COHORT
Time perspective
RETROSPECTIVE
Eligibility
Sex/Gender
ALL
Age
18 Years to No maximum
Healthy volunteers
No
Inclusion criteria
Adults who underwent colon resection surgery at an AP-HP hospital between 01/01/2017 and 01/11/2024, with: A preoperative abdominopelvic CT scan available within 60 days prior to surgery. A corresponding pathology report (anatomopathological results) available within 90 days post-surgery. Colon resection identified by CCAM procedure codes: HHFA002, HHFA004, HHFA005, HHFA006, HHFA008, HHFA009, HHFA010, HHFA014, HHFA017, HHFA018, HHFA021, HHFA022, HHFA023, HHFA024, HHFA026, HHFA028, HHFA029, HHFA030, HHFA031, HHFC040, HHFC296. Confirmed diagnosis of colon tumor by ICD-10 code: C18\* (colonic neoplasms).
Exclusion criteria
Patients who received neoadjuvant chemotherapy prior to surgery, identified by ICD-10 codes Z511 or Z512 recorded before the surgical act. These exclusions will be refined and confirmed through manual medical record review to ensure accuracy. Absence of usable CT imaging or anatomical pathology data linked to the surgical event.
Design outcomes
Primary
| Measure | Time frame | Description |
|---|---|---|
| Diagnostic performance of CT-based deep learning models for T (T1-2 vs T3-4) and N (N- vs N+) staging. | Index preoperative CT through postoperative pathology report (within 90 days of surgery). | Area under the ROC curve (AUC) and F1-score of deep learning models in (a) classifying early vs advanced T stage (T1-2 vs T3-4) and (b) nodal status (N- vs N+) compared with pathology gold standard. Sensitivity, specificity, PPV, NPV reported as supportive metrics; model performance compared to historical radiologist benchmarks when available. |
Secondary
| Measure | Time frame | Description |
|---|---|---|
| Detection performance for T4 tumors on preoperative CT. | Index CT to pathology confirmation (≤90 days post-surgery). | AUC, F1, sensitivity, specificity for binary classification T4 vs non-T4 (T1-3). Analyses stratified by tumor location and contrast phase when data permit. |
| Multiclass T-stage classification accuracy (T1, T2, T3, T4). | Index CT to pathology confirmation (≤90 days). | Macro-averaged F1, per-class sensitivity/specificity, confusion matrix, and Cohen's kappa comparing model-predicted 4-class T stage with pathology. |
| Prognostic value of CT-derived model features for clinical outcomes and survival. | From index CT to last follow-up (up to 5 years, or maximum available follow-up in EHR). | Association between model outputs (probabilities, embeddings) and (a)coverall survival, (b) disease-free survival. Evaluated using Kaplan-Meier analysis, log-rank tests, and multivariable Cox models adjusted for clinical covariates where available. |
Countries
France
Contacts
CONTACTQuentin Vanderbecq, MD
CONTACTMathilde WAGNER, MD,PhD
PRINCIPAL_INVESTIGATORQuentin Vanderbecq, MD
Assistance Publique - Hôpitaux de Paris
Outcome results
None listed