Leveraging Multi-Phase CT for the Automatic Diagnosis of HCC with Deep Learning: Impacts of Lesion Localization and Phase Inclusion

The early and accurate diagnosis of hepatocellular carcinoma is paramount to allow the successful treatment of this deadly liver cancer, of rising incidence worldwide. With the ultimate objective to design a deep learning method able to overcome the limitations of the LI-RADS scoring system, this study systematically and quantitatively evaluates how key parameters impact the diagnostic performance of a simple classification network. Taking advantage of a diverse database of multiphase abdominal CT scans, various approaches for liver lesion localization and phase inclusion were benchmarked, as well as coping strategies for residual misalignments and missing phases. A competitive AUC of 90% was achieved when classifying quadri-phase input patches centered around the center of gravity of venous-phase lesion segmentation maps. Youden’s sensitivity, specificity and balanced accuracy were 82%, 90% and 86%. This performance degraded when cross-phase residual misalignments were simulated or when lesser phases were exploited, but data augmentation and data imputation approaches allowed to reduce the drop in performance.