FRODO: An In-Depth Analysis of a System to Reject Outlier Samples From a Trained Neural Network

Çallı, Erdi; Ginneken, Bram van; Sogancioglu, Ecem

doi:10.1109/tmi.2022.3221898

Cited by 4 publications

(9 citation statements)

References 15 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…21 Instead of converting the model's output to an ID score, several approaches use the activations of the model to generate class-conditional Gaussian distributions (Mahalanobis). 24,25 This method models OOD images as unlikely points in the class distribution, that is, having a large Mahalanobis distance to the modeled class means in the latent space. Lee et al 25 motivated the use of the Mahalanobis distance between the mean representation of a class and the input in the feature space, instead of performing OOD detection in the label space due to "label overfitting", that is, that the model predictions are conditioned on the training labels.…”

Section: Other Ood Detection Methodsmentioning

confidence: 99%

“…Lee et al 25 motivated the use of the Mahalanobis distance between the mean representation of a class and the input in the feature space, instead of performing OOD detection in the label space due to "label overfitting", that is, that the model predictions are conditioned on the training labels. For Mahalanobis-based OOD detection, we use the output of the penultimate layer to determine the Mahalanobis scores similar to the work by Çallı et al 24 Hendrycks et al propose training the classification model with additional self -supervised heads to improve OOD robustness (SS OOD). 23 In this method, the model has to additionally predict image rotation and translation.…”

Section: Other Ood Detection Methodsmentioning

confidence: 99%

“…[16][17][18] Therefore, in recent years, several methods, have been proposed to detect OOD samples. [19][20][21][22][23][24] Commonly, the OOD detector converts the output of a model to an ID score. For example, one of the earliest approaches, Max.…”

Section: Introductionmentioning

confidence: 99%

“…In their experiments, the authors noticed lower confidence scores for the highest class probability for OOD inputs compared to ID inputs. Another approach, proposed by Lee et al, 25 also applied to chest x-rays, 24 models OOD data based on the smallest Mahalanobis distance between the input and a class conditional Gaussian distribution in the latent space. Furthermore, Hendrycks et al propose to use a selfsupervised training scheme to improve OOD detection performance.…”

Section: Introductionmentioning

confidence: 99%

“…So far, the problem caused by OOD data has been investigated mostly on toy data sets, for example, a model trained on the CIFAR-10 data set, 27 learning to classify automobiles and trucks, is tested on the SVHN data set 28 containing house numbers,or in-house data sets. 24 This raises the question if the test performance of proposed OOD detectors translate to an existing model trained on chest x-rays. Figure 1 motivates this problem: as the real world data consists of more than frontal chest x-rays, a classifier like CheXnet 3 must handle OOD images safely.…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Out‐of‐distribution detection with in‐distribution voting using the medical example of chest x‐ray classification

Wollek,

Willem,

Ingrisch

et al. 2023

Medical Physics

View full text Add to dashboard Cite

BackgroundDeep learning models are being applied to more and more use cases with astonishing success stories, but how do they perform in the real world? Models are typically tested on specific cleaned data sets, but when deployed in the real world, the model will encounter unexpected, out‐of‐distribution (OOD) data.PurposeTo investigate the impact of OOD radiographs on existing chest x‐ray classification models and to increase their robustness against OOD data.MethodsThe study employed the commonly used chest x‐ray classification model, CheXnet, trained on the chest x‐ray 14 data set, and tested its robustness against OOD data using three public radiography data sets: IRMA, Bone Age, and MURA, and the ImageNet data set. To detect OOD data for multi‐label classification, we proposed in‐distribution voting (IDV). The OOD detection performance is measured across data sets using the area under the receiver operating characteristic curve (AUC) analysis and compared with Mahalanobis‐based OOD detection, MaxLogit, MaxEnergy, self‐supervised OOD detection (SS OOD), and CutMix.ResultsWithout additional OOD detection, the chest x‐ray classifier failed to discard any OOD images, with an AUC of 0.5. The proposed IDV approach trained on ID (chest x‐ray 14) and OOD data (IRMA and ImageNet) achieved, on average, 0.999 OOD AUC across the three data sets, surpassing all other OOD detection methods. Mahalanobis‐based OOD detection achieved an average OOD detection AUC of 0.982. IDV trained solely with a few thousand ImageNet images had an AUC 0.913, which was considerably higher than MaxLogit (0.726), MaxEnergy (0.724), SS OOD (0.476), and CutMix (0.376).ConclusionsThe performance of all tested OOD detection methods did not translate well to radiography data sets, except Mahalanobis‐based OOD detection and the proposed IDV method. Consequently, training solely on ID data led to incorrect classification of OOD images as ID, resulting in increased false positive rates. IDV substantially improved the model's ID classification performance, even when trained with data that will not occur in the intended use case or test set (ImageNet), without additional inference overhead or performance decrease in the target classification. The corresponding code is available at https://gitlab.lrz.de/IP/a‐knee‐cannot‐have‐lung‐disease.

show abstract

Section: Other Ood Detection Methodsmentioning

confidence: 99%

Section: Other Ood Detection Methodsmentioning

confidence: 99%