Impact of DWI quality and tumor complexity on deep learning-based segmentation of rectal cancer using multicenter MRI data

Abstract Objectives To assess the influence of DWI quality and tumor complexity on the performance of a fully automated deep learning segmentation algorithm in a large and heterogeneous clinical dataset of rectal cancer MRIs. Materials and methods In this retrospective multicenter study, nnU-Nets were trained on baseline staging MRIs using either a dataset of T2W-MRI, DWI, and ADC maps, or only T2W-MRI. Scans were randomly split into a train and test cohort (ratio 8:2). Preprocessing, hyperoptimization, and training were performed in the nnU-Net framework using 5-fold cross-validation on the training set. An expert-radiologist manually delineated all tumors on high b-value DWI, with T2W-MRI for anatomical correlation to serve as training input and as the ground truth to test network performance using the Dice similarity coefficient (DSC). A second expert-radiologist independently segmented all patients from the test cohort to calculate network performance and expert readers’ agreement. DWI quality and morphological tumor complexity were assessed using 3–5 point Likert scores to evaluate their respective effects. Results 603 rectal cancer patients (mean age, 65 ± 19.1 years, 376 men, scanned in 6 institutions from 2012 to 2017) were analyzed. Using the combined T2W/DWI/ADC dataset, the network showed good performance (DSC 0.669–0.726 in the test set), results comparable to expert readers’ agreement (DSC 0.757). Results based on T2W data only were significantly poorer (DSC 0.586–0.590). More complex tumors resulted in lower network performance and expert readers’ agreement. DWI quality had no significant impact. Conclusion Despite large heterogeneity in data, the network performed comparable to expert readers, hampered only by tumor complexity. Key Points Question Deep learning models for tumor segmentation have predominantly been trained and tested in highly curated datasets and require further validation in real-world clinical data. Findings Deep learning-based segmentation achieved performance comparable to expert reader agreement; increased tumor complexity reduced segmentation performance, whereas DWI quality had no significant impact. Relevance statement Despite heterogeneous data, deep learning can provide good-quality segmentations that can be used as a starting point for further research and to ultimately reduce segmentation workloads.