MRI Radiomics Model Predicts Pathologic Complete Response of Rectal Cancer Following Chemoradiotherapy.

Background Preoperative assessment of pathologic complete response (pCR) in locally advanced rectal cancer (LARC) after neoadjuvant chemoradiotherapy (nCRT) is increasingly needed for organ preservation, but large-scale validation of an MRI radiomics model remains lacking. Purpose To evaluate radiomics models based on T2-weighted imaging and diffusion-weighted MRI for predicting pCR after nCRT in LARC and compare their performance with visual assessment by radiologists. Materials and Methods This retrospective study included patients with LARC (clinical stage T3 or higher, positive nodal status, or both) who underwent post-nCRT MRI and elective resection between January 2009 and December 2018. Surgical histopathologic analysis was the reference standard for pCR. Radiomic features were extracted from the volume of interest on T2-weighted images and apparent diffusion coefficient (ADC) maps from post-nCRT MRI to generate three models: T2 weighted, ADC, and both T2 weighted and ADC (merged). Radiomics signatures were generated using the least absolute shrinkage and selection operator with tenfold cross-validation. Three experienced radiologists independently rated tumor regression grades at MRI and compared these with the radiomics models' diagnostic outcomes. Areas under the curve (AUCs) of the radiomics models and pooled readers were compared by using the DeLong method. Results Among 898 patients, 189 (21%) achieved pCR. The patients were chronologically divided into training ( = 592; mean age ± standard deviation, 59 years ± 12; 388 men) and test ( = 306; mean age, 59 years ± 12; 190 men) sets. The radiomics signatures of the T2-weighted, ADC, and merged models demonstrated AUCs of 0.82, 0.79, and 0.82, respectively, with no evidence of a difference found between the T2-weighted and merged models ( = .49), while the ADC model performed worse than the merged model ( = .02). The T2-weighted model had higher classification performance (AUC, 0.82 vs 0.74 [ = .009]) and sensitivity (80.0% vs 15.6% [ < .001]), but lower specificity (68.4% vs 98.6% [ < .001]) than the pooled performance of the three radiologists. Conclusion An MRI-based radiomics model showed better classification performance than experienced radiologists for diagnosing pathologic complete response in patients with locally advanced rectal cancer after neoadjuvant chemoradiotherapy. © RSNA, 2022 See also the editorial by Taylor in this issue.