Framework for Assessing the Impact of CNN-based Image Segmentation on Multi-step Biomarker Extraction

Convolutional neural networks (CNN) have become endemic to medical image analysis. However, little work has assessed the effect of CNN error on multi-step biomarker extraction. Here, we describe a framework for assessing CNN impact on final biomarker extraction in the clinical application of detecting colitis on ¹⁸F-FDG PET/CT.

We employ a CNN to segment the bowel on CT, quantify inflammation using ¹⁸F-FDG PET histogram metrics, and perform sensitivity analysis to optimize the extracted metric for our clinical task. CNN performance is characterized by Dice similarity coefficient (DSC). Perturbation-based analysis is used to quantify the impact of segmentation error on PET histogram metrics. Area under the receiver operating characteristic curve (AUC) is used to optimize the PET histogram metric for our clinical task.

The CNN had a validation DSC of 0.87±0.06 (mean±sd) for bowel segmentation. PET histogram metrics were robust to small dilations, erosions, and spatial shifts. The 96^th percentile of the PET histogram was determined to be the optimal biomarker for classifying patients with an AUC of 0.91.

The presented framework is a method by which CNN segmentation error can be accounted for in medical image analyses.