Current limitations to identify covid-19 using artificial intelligence with chest x-ray imaging (part ii). The shortcut learning problem

Abstract

Since the outbreak of the COVID-19 pandemic, computer vision researchers have been working on automatic identification of this disease using radiological images. The results achieved by automatic classification methods far exceed those of human specialists, with sensitivity as high as 100% being reported. However, prestigious radiology societies have stated that the use of this type of imaging alone is not recommended as a diagnostic method. According to some experts the patterns presented in these images are unspecific and subtle, overlapping with other viral pneumonias. This report seeks to evaluate the analysis the robustness and generalizability of different approaches using artificial intelligence, deep learning and computer vision to identify COVID-19 using chest X-rays images. We also seek to alert researchers and reviewers to the issue of “shortcut learning”. Recommendations are presented to identify whether COVID-19 automatic classification models are being affected by shortcut learning. Firstly, papers using explainable artificial intelligence methods are reviewed. The results of applying external validation sets are evaluated to determine the generalizability of these methods. Finally, studies that apply traditional computer vision methods to perform the same task are considered. It is evident that using the whole chest X-Ray image or the bounding box of the lungs, the image regions that contribute most to the classification appear outside of the lung region, something that is not likely possible. In addition, although the investigations that evaluated their models on data sets external to the training set, the effectiveness of these models decreased significantly, it may provide a more realistic representation as how the model will perform in the clinic. The results indicate that, so far, the existing models often involve shortcut learning, which makes their use less appropriate in the clinical setting.

Development and Validation of a Multimodal-Based Prognosis and Intervention Prediction Model for COVID-19 Patients in a Multicenter Cohort

“…By separating the training and validation data from each center, we were able to show that our multimodal approach achieved high performance and generalized well to data that are characteristically different from the training dataset and have not been seen before. It has been shown that previous imaging-based AI models for COVID-19 prognostication may be inadequate for real clinical usage [35,36]. The major reason for this was the lack of external validation data.…”

POTHER: Patch-Voted Deep Learning-Based Chest X-ray Bias Analysis for COVID-19 Detection

“…Most of the approaches developed for the classification of the CXR images relay on global features [2], [11], [12], [14]. However, these features may not accurately represent the complex nature of CXR images [22]. It should be verified when working with medical images, especially when the model's accuracy is very high [13], [21], the reasons why algorithms perform so well to prevent developing algorithms which base their decision on confounding factors rather then medical pathology.…”

“…They create a map of lung areas important for COVID-19 detection and claim a production-ready solution. Considering the need for a critical approach to chest classification and analysis of biases [23], [22], [20], we carry out further research into the analysis of the deep learning model decisions using the COVIDx dataset.…”

“…We use XAI methods to demonstrate that model decisions may rely on confounding factors rather than medical pathology. Degrave et al [20] and Cabrera et al [22] call them shortcuts, while Maguolo et al [23] biases. An analysis of shortcuts in CXRs, which we call confounding biases (CBs), based on open-source datasets and global learning methods, is presented by [20].…”

Generalizability of CNN on Predicting COVID-19 from Chest X-ray Images

“…In one hand, there may be visible lesions on CT that are not visible on CXR images. On the other, among CXR and CT scans, CXR images are interesting because they have a lower associated cost, are faster to acquire, and are more widely available [Pereira et al 2020], [López-Cabrera et al 2021]. In this work, we focus our attention on the CXR images.…”