1352042 - Detecting AI failure in CMR segmentation at slice level: towards easy and robust AI quality control

Despite the impressive performance of deep learning in cardiac magnetic resonance imaging (CMR) segmentation, well-trained neural networks may still fail at test time [1]. Monte Carlo (MC) Dropout [2] can provide pixel-level uncertainty estimation, however, it is reported to cause “silent failure” (i.e., fail to detect failure) [3]. Additionally, pixel-level uncertainty demands intensive visual inspection by radiologists/cardiologists. This study investigates the feasibility of detecting AI failure at slice level using our recently proposed Bayesian deep learning by Checkpoint Ensemble [4]. The purpose is to enable easy and robust AI quality control in clinic. 

Methods:

The Bayesian segmentation network by MC-Dropout and Checkpoint Ensemble were trained on the public ACDC dataset [5] with 100 short-axis cine steady state free precession (SSFP) CMR scans acquired from Siemens Area 1.5T and Siemens Trio Trim 3.0T and tested on the M&M challenge dataset [6] which consists of cine SSFP scans of 150 patients acquired from Siemens MAGNETOM Avanto 1.5T and Philips Achieva 1.5T. Using manual annotation as ground-truth, we designate it as AI failure if on a particular slice, the dice coefficient is below 85% for left/right ventricle (LV/RV) and 80% for myocardium, and the 95-percentile Hausdorff distance exceeds 3 mm. We derive 3 segmentation confidence scores, to indicate success or failure at slice level for LV, RV, and myocardium respectively, via averaging pixel-level uncertainty of the predicted region for each cardiac structure. We evaluated the proposed segmentation confidence score by the area under curve (AUC) of the receiver operating characteristic (ROC).

Results:

Table 1 reports the AUC values, showing that the segmentation confidence score from both Bayesian methods can effectively detect segmentation failure on LV (93.4% vs. 93.8%) and myocardium (82.6% vs. 84.3%). The segmentation confidence score based on our Checkpoint Ensemble better detected segmentation failure on RV than MC-Dropout (AUC 91.6% vs. 87.9%). Figure 1 (a) – (c) shows three examples of segmentation results and the corresponding RV segmentation confidence score by MC-Dropout (first row) and Checkpoint Ensemble (second row). In all three cases, RV segmentation failed, but the failure was not detected by the segmentation confidence score from MC-Dropout (“silent failure”). In contrast, all three failure cases were detected by the segmentation confidence score based on Checkpoint Ensemble.

Conclusion:

We proposed and validated a segmentation confidence score to detect AI CMR segmentation failure at slice level. The confidence score based on our Bayesian Checkpoint Ensemble demonstrated higher sensitivity than traditional MC-Dropout, especially on RV which is more challenging to segment. High sensitivity in failure detection prevents silent failure and potentially translates to reliable AI in clinical practice.