Abstract 546P
Background
More than 70% of patients with T1 colorectal cancer (CRC) undergo a radical surgery involving lymph node dissection, despite the presence of lymph node metastasis (LNM) being observed in only ∼10% of cases. To reduce unnecessary radical surgery, we developed a multimodal artificial intelligence (AI) system to identify T1 CRC at risk for LNM.
Methods
We analyzed 402 CRC patients from two independent multicenter cohorts, including a training (n = 219) and a validation cohort (n = 183). Proteomics dataset from tumor tissue were identified and quantified using LC-MS/MS analysis. Multiple pathomics features were extracted from whole H&E slides with patch-level convolutional neural network training in weakly supervised manner. Clinicopathological characteristics included age, tumor size, location, lymphatic and vascular invasion, perineural invasion, histologic grade, and CEA. A machine-learning artificial neural network using proteomics, pathomics and clinicopathological features was developed.
Results
LNM were detected in 78 out of 219 patients (35.6%) within the training cohort and in 29 out of 183 patients (15.8%) within the validation cohort. A panel of six protein biomarkers (OSBPL5, ATAD2, BAIAP2, MANBA, ITPR2, ARHGAP5) screened by proteomics was quantified by immunohistochemistry and subsequently integrated into the AI system. In the validation cohort, the multimodal AI system achieved robust classification of LNM status, yielding an AUC of 0.978. In comparison, the guideline-based model exhibited an AUC of 0.717 for identifying LNM (P < 0.001).
Conclusions
The AI system integrating deep learning-proteomics, pathomics and clinicopathological features robustly identifies T1 CRC patients at risk of LNM in a preoperative setting. The multimodal AI system would improve clinical practice by alleviating the burden of unnecessary overtreatment for T1 CRC.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
Jianmin Xu.
Funding
The National Natural Science Foundation of China.
Disclosure
All authors have declared no conflicts of interest.
Resources from the same session
523P - A phase II study of intrapatient dose escalation of biweekly trifluridine/tipiracil plus bevacizumab for colorectal cancer (E-BiTS study)
Presenter: Munehiro Wakabayashi
Session: Poster session 15
524P - A phase II study of alpelisib, a PIK3CA inhibitor, and capecitabine in patients with metastatic colorectal cancer who failed two prior standard chemotherapies
Presenter: Ahreum Lim
Session: Poster session 15
525P - Final analysis of the JACCRO CC-16: Ramucirumab plus FOLFIRI for RAS wild-type mCRC refractory to anti-EGFR antibody
Presenter: Keiji Sugiyama
Session: Poster session 15
526P - Efficacy and safety of fruquintinib in refractory metastatic colorectal cancer: A FRESCO-2 subgroup analysis by age
Presenter: Maria Elena Elez Fernandez
Session: Poster session 15
527P - Effect of trifluridine/tipiracil (FTD/TPI) in combination with bevacizumab (bev) in patients treated in SUNLIGHT by clinical prognostic factors at baseline
Presenter: Josep Tabernero
Session: Poster session 15
529P - Evaluating metastatic disease sites as a prognostic marker in patients receiving sequential treatment with regorafenib and trifluridine/tipiracil for refractory colorectal cancer: Survival outcomes from the multicenter retrospective “ReTrITA” study
Presenter: Carlo Signorelli
Session: Poster session 15
530P - Real-world effectiveness and predictive biomarker analysis of TAS-102+bevacizumab vs. regorafenib vs. TAS-102 in metastatic colorectal cancer: A multicenter cohort study
Presenter: Andreas Seeber
Session: Poster session 15
531P - Prognostic value of liver metastases, KRAS mutations and race in colorectal cancer patients: A pooled analysis of third-line placebo-controlled trials from the ARCAD database
Presenter: Thierry André
Session: Poster session 15