Clinical and Radiomics-Based Deep Learning Predictive Model for Early Treatment Failure after Neoadjuvant Radiochemotherapy for Rectal Cancer

Purpose/Objective(s)

As shown in RAPIDO and PRODIGE-23 trials, total neoadjuvant therapy for rectal cancer increases the complete response rate, however, the overall survival benefit remains uncertain. Although postponing surgery assures more time for tumor downstaging, it also increases surgical complications and can worsen prognosis in non-responding tumors. We aimed to develop a clinical and radiomics-based model for early treatment failure (ETF) prediction after neoadjuvant radiochemotherapy (RCHTx) for rectal cancer.

Materials/Methods

We conducted a prospective cohort analysis of 92 consecutive and eligible patients with rectal adenocarcinoma treated with neoadjuvant radiochemotherapy in the years 2018-2020, and followed until February 2022. We defined ETF as postoperative poor pathological response and no MRI-confirmed downstaging; or unresectability after neoadjuvant treatment; or cancer-related or treatment-related death within 18 months since treatment initiation. Collected clinical data included physical examination, grading, staging (before and after RCHTx), biomarker levels, RCHTx prescription details, and pre-operative MRI tumor and nodal characteristics. Radiomic features were extracted from GTV delineated on axial MRI T2 TSE sequences. To develop a predictive model, we followed OmicSelector-based (https://biostat.umed.pl/OmicSelector) feature selection, logistic regression (LR), and artificial deep neural network model development procedure. After splitting into training (60%), testing (20%), and validation (20%) datasets, the final feature set was selected based on the highest average logistic regression accuracy on all sets, while all patients with MRI performed outside our institution were included in the testing set (compensating possible batch effect).

Results

ETF was experienced by 24% of patients and was associated with significantly shorter overall survival (median 12.1 months [95%CI:8.5-16.5] vs. not reached, p<0.0001). None of the clinical factors was significantly associated with ETF (adjusted p value>0.05; AUC ROC range 0.39-0.64). LR model based on clinical factors presented overfitting (accuracy 82% on the training set, 56.2% on testing and 68.8% on validation set). The best feature set included 14 selected radiomic features, patient's weight, T-stage, grading, mesorectal fascia and nodal involvement. LR model based on those features achieved perfect accuracy on training and testing sets and 81.3% accuracy on the validation set. Deep neural network model utilizing selected clinical and radiomic features achieved 87.5% accuracy (80% sensitivity, 91% specificity) on testing and 93.8% accuracy (100% sensitivity, 91% specificity) on validation set.

Conclusion

Combined clinical and radiomics-based deep learning model could predict ETF after RCHTx for rectal cancer with at least 80% sensitivity and 90% specificity. Prediction of ETF after RCHTx could select high-risk patients who may require non-standard treatment.