551P - Efficacy prediction of chemoradiotherapy plus anti-PD-1/PD-L1 treatment by magnetic resonance imaging in MSS locally advanced rectal cancer

Background

PD-1/PD-L1 blockades with chemoradiotherapy as neoadjuvant therapy may promote tumor regression in patients with microsatellite stable (MSS) rectal cancer. However, which patient can benefit from this new treatment remains largely unknown, especially through analysis of magnetic resonance imaging (MRI) images.

Methods

Two phase II, open-labeled, studies evaluated the efficacy of anti-PD-1/PD-L1 to neoadjuvant chemoradiotherapy in MSS rectal cancer. Patients with locally advanced middle and low rectal cancer were recruited. All patients received long-course radiotherapy followed by two or three cycles anti-PD-1/PD-L1 treatment and radical resection surgery. Rectal MRI images were collected from patients prior to the radiotherapy. From each patient's rectal MRI set, three images where the tumor was clearly visible were selected and labeled. A convolutional neural network, named PDNet, was developed to extract features from these selected images and predict whether a pathological complete regression (pCR) is achieved.

Results

61 patients were recruited with 100% R0 resection. Among them, 24 patients achieved pCR (39.3%). A dataset comprising 183 rectal MRI images was utilized, with 147 images employed in the training phase and 36 images in the testing phase. After training, the PDNet model demonstrated strong performance on the test dataset (internal validations): it achieved a classification accuracy of 86.11%, a positive predictive value of 93.33%, a negative predictive value of 80.95%, a sensitivity of 77.78%, and a specificity of 94.44%. The model also recorded a p-value of 0.0000498 and an Area Under the Curve (AUC) of 0.8580, with a 95% Confidence Interval (CI) [0.6982, 0.9805].

Conclusions

The study concludes that features extracted from MRI images are highly correlated with the likelihood of achieving pCR. Using the developed neural network model (PDNet), we can predict the probability of achieving pCR before the treatment of PD-1/PD-L1 blockades with chemoradiotherapy in MSS local advanced rectal cancer. External validation of the model is currently underway.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

Jianmin Xu.

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.