Development and evaluation of a deep learning framework for detection and diagnosis of peri-ampullary tumor in MRI images

Abstract

Abstract Purpose: We were development and evaluation of one deep learning (DL) framework for identification of Peri-ampullary (PA) regions and diagnosis of peri-ampullary tumor (PAT) conditions in MRI images. Experimental Design: We retrospectively analyzed 1,038 patients. We found that DL algorithm for identification of PA regions and diagnosis of PAT conditions in MRI images. Results: The DL algorithm successfully identified and segmented the PA regions in both T1WI (IOU = 0·62) and T2WI images (IOU = 0·55). Based on the segmentations of PA regions in images, the classification DL algorithm achieved optimal accuracies in classifications of NPA and PSOL with AUC of 0·71 (95% CI 0·68 to 0·74) (T1WI) and 0·72 (95% CI 0·68 to 0·75) (T2WI). For PSOL cases, another classification DL algorithm achieved encouraging accuracies to further classify PAT and Non-PATL with AUC of 0·81 (95% CI 0·77 to 0·85) (T1WI) and 0·78 (95% CI 0·73 to 0·83) (T2WI). Furthermore, in the patient-based approach, the classification DL algorithm achieved optimal accuracies in classifications of NPA and PSOL with ACC of 0·75 (95% CI 0·65 to 0·85) (T1WI) and 0·88 (95% CI 0·81 to 0·94) (T2WI). For PSOL cases, another classification DL algorithm achieved encouraging accuracies to further classify PAT and Non-PATL with ACC of 0·83 (95% CI 0·71 to 0·96) (T1WI) and 0·82 (95% CI 0·70 to 0·93) (T2WI). Conclusions: Our study suggests that DL could accurately identify and segment PA regions in MRI images and further classify PAT conditions with promising accuracies. DL could assist clinicians in MRI interpretation for PAT diagnosis.