Phase recognition in contrast-enhanced CT scans based on deep learning and random sampling

Abstract

Purpose: A fully automated system for interpreting abdominal computed tomography (CT) scans with multiple phases of contrast enhancement requires accurate classification of the phases. Current approaches typically utilize three-dimensional (3D) convolutional neural networks (CNNs), which are often complex and have high latency. This work aims to develop and validate a precise, fast multi-phase classifier for recognizing the main types of contrast phases in abdominal CT scans.

Methods: This study proposes a novel method that employs a random sampling mechanism in conjunction with deep CNNs for phase recognition of abdominal CT scans across four phases: non-contrast, arterial, venous, and others. The CNNs perform slice-wise phase predictions, while random sampling selects input slices for the CNN models. Majority voting synthesizes the slice-wise results to provide a final prediction at the scan level.

Results: Our classifier was trained on 271,426 slices from 830 phase-annotated CT scans. By combining majority voting on 30% of randomly chosen slices from each scan, we achieved a mean F1 score of 92.09% on our internal test set of 358 scans. The method was also evaluated on two external test sets: CTPAC-CCRCC (N = 242) and LiTS (N = 131), yielding mean F1 scores of 76.79% and 86.94%, respectively. Despite a drop in performance, the model maintained high accuracy and significantly outperformed existing 3D approaches while requiring less computation time for inference.

Conclusions: The proposed approach demonstrates better accuracy and significantly reduced latency compared to state-of-the-art classification methods. This study showcases the potential of a precise, fast multiphase classifier based on a two-dimensional deep learning approach combined with random sampling for contrast phase recognition, offering a valuable tool for extracting multiphase abdominal studies from low veracity, real-world data.