Deep learning for automatic detection of prostatectomy in CT images of prostate cancer patients.

Abstract

Extracting anatomic reference from computed tomography (CT) images is a crucial step in fully automated image analysis solutions for CT and hybrid imaging of prostate cancer. Several deep learning-based applications can be used to perform segmentation of different anatomical structures in CT, but they might produce a false prostate segment for post-treatment scans of patients treated with prostatectomy. In this study, our aim was to both systematically assess the performance of a state-of-the-art CT segmentation tool, TotalSegmentator, for post-prostatectomy patients and to investigate the potential of using a convolutional neural network (CNN) to automatically detect prostatectomy from CT images. We collected a dataset of CT images from 542 patients, 269 of which were treated with robotic-assisted laparoscopic radical prostatectomy (RP), 194 of which were treated with radiation and/or androgen deprivation therapy only, and 79 of which were treatment-naive. We used TotalSegmentator to perform multi-organ segmentation for all the patients, computed the volumes of the greatest connected components of the prostate segments, and studied the use of a cut-off threshold for the resulting volumes to detect RP with five-fold cross-validation. Additionally, we trained and evaluated a light-weight CNN for classifying patients treated with and without RP. According to our results, TotalSegmentator produced a false prostate segment for 98.5% patients treated with RP. The prostate segments were significantly smaller in the RP cohort than the other two cohorts (12.1 ± 5.8 cm[Formula: see text] vs. 22.8 ± 8.9 cm[Formula: see text] and 22.4 ± 10.9 cm[Formula: see text], p-values: 1.3e-17 and 2.1e-17). The use of cut-off thresholds for TotalSegmentator's prostate volumes resulted in RP detection accuracy of 77.5 ± 1.3%, sensitivity of 83.0 ± 3.9%, specificity of 72.2 ± 3.5%, precision of 74.4 ± 3.1%, and area under receiver operating characteristic curve (auROC) of 84.1 ± 2.3%. Conversely, our proposed CNN approach detected RP with accuracy of 86.5 ± 4.2%, sensitivity of 78.9 ± 12.5%, specificity of 93.5 ± 3.9%, precision of 92.7 ± 3.9%, and auROC of 95.1 ± 1.7%, outperforming the threshold approach in a statistically significant way according to DeLong's tests (4 out of 5 p-values<0.045). To conclude, TotalSegmentator systematically produces false prostate segments with patients treated with RP but, by utilizing a CNN, RP can be detected automatically based on CT data only. This enables automated correction of TotalSegmentator masks. If developed further, our CNN method could provide a stand-alone application for identifying presence of prostatectomy efficiently without the need to consult treatment records and, consequently, assist in developing fully-automated image analysis tools.