Key points are not available for this paper at this time.
Accurate identification of pathological complete response (pCR) following neoadjuvant chemoradiotherapy for locally advanced rectal cancer (LARC) is critical for non-operative "watch-and-wait" strategies. In this study, we evaluated whether co-registered acoustic resolution photoacoustic microscopy and ultrasound (ARPAM-US) endoscopy with deep learning could better identify pCR than T2-weighted MRI radiomics models. A pretrained ResNet50 model utilizing ARPAM-US B-scans was developed on a prospective cohort (n = 25) to predict pCR, based on surgical histopathological assessment, using 5-fold cross-validation. MRI radiomics models were trained on a retrospective cohort (n = 119). Performances between ARPAM-US ResNet50 and MRI radiomics models were compared. The ARPAM-US model achieved promising diagnostic performance (AUC 0.956, 95% CI 0.912-1.000), distinguishing normalized vascular architecture in complete responders. MRI radiomics models demonstrated degraded prospective generalization. While limited responders precluded definitive head-to-head comparison, deep learning-enhanced ARPAM-US endoscopy demonstrates promising diagnostic accuracy for assessing rectal cancer treatment response. Larger prospective studies are needed to validate these findings.
Nie et al. (Sat,) studied this question.