Patient body fat and AI ensemble technique and the influence on false-positive rate in AI second observer for colorectal cancer detection.

Abstract

216 Background: Colorectal cancer (CRC) is the second leading cause of cancer-related deaths. We previously found that delayed diagnosis due to lack of radiological identification results in significantly worse outcome for patients. We had developed a rudimentary, AI second observer which demonstrated potential for detecting CRC on routine CT abdomen/pelvis (CTAP). However, the AI algorithm detected many false positives. In this study, we analyzed the data using TCIA as test cases and evaluated whether patient peritoneal fat content influenced the false positive rate. This could serve as a guide for future training of AI second observer to minimize false positive detection. Methods: 2D U-Net convolutional neural network (CNN) containing 31 million trainable parameters was trained with 58 CRC CT images from Banner MD Anderson (AZ) and MD Anderson Cancer Center (TX) (51 used for training and 7 for validation) and 59 normal CT scans from Banner MD Anderson Cancer Center. 18 of the 25 CRC cases from public domain data (The Cancer Genome Atlas) were used to evaluate the performance of the models (5 had no identifiable cancer and 2 were rejected for having no contrast). The CRC was segmented using ITK-SNAP open-source software (v. 3.8). To apply the deep ensemble approach, five CNN models were trained independently with random initialization using the same U-Net architect and the same training data. Given a testing CT scan, each of the five trained CNN models was applied to produce tumor segmentation for the testing CT scan. The tumor segmentation results produced by the trained CNN models were then fused using a simple majority voting rule (up to 2 voters) to produce consensus tumor segmentation results. The segmentation was analyzed for the number and location of false positives per case. The peritoneal fat content was classified at the level of aortic bifurcation by the distance of fat between adjacent small bowel loops (≤ or > 1 cm). Chi-square test was performed testing fat volume and number of voters as the intervention. Results: Our results showed that the higher volume of peritoneal fat (> 1 cm, N=6) decreases the rate of false positive compared with low volume (≤ 1 cm, N=12, p=0.013). When comparing between having one voter and two voter ensemble using low fat volume data, two voter ensemble also decreased the number of false positives but not statistically significant (p=0.286). Conclusions: Our results show that AI-based second observer generates more false positives when patients have lower peritoneal fat volume; this implies that future training may require higher percentage of cases with low peritoneal fat to improve second observer precision. Our analysis also showed that increasing the number of voter in the ensemble also decrease the number of false positives per case.